You are on page 1of 61

# 56

CHAPTER 2 Review of Probability 2.5 Suppose that K , . . . ,Yn are i.i.d. random variables with a N(l. 4) distribution. Sketch the probability density of Y when n — 2. Repeat this for n = 10 and n = 100. In words, describe how the densities differ. What is the relationship between your answer and the law of large numbers? Suppose that Y\,...,Yn are i.i.d. random variables with the probability distribution given in Figure 2.10a. You want to calculate Pr( Y < 0.1). Would it be reasonable to use the normal approximation if n -- 5? What about n = 25 or n - 100? Explain. Y is a random variable with ixY = 0,aY— 1, skewness - 0, and kurtosis = 100. Sketch a hypothetical probability distribution of Y. Explain why n random variables drawn from this distribution might have some large outliers. .«&<• Exercises Non-college grads (X = 0) College grads(A'=l) Total 0i. V

Exercises

57

«

S

M

Unemployed <Y= 0) 0 037 0 009 0 046

Employed <K= 1) 0.622 0.332 0.954

Total 0.659 0.341 1.000

2.6

2.7

a. Compute E(Y).

0^5^

: ^ ^ jfit j* b. The unemployment rate is the fraction of the labor force that is ^ '<^ unemployed. Show that the unemployment raje is given by 1 E{ Y)

a.

yS\ *\\s&< <?stJc O 1+ C& * V\* V4.J ^\< • \ i* y X U 7*

ttf

Let Y denote the number of "heads" that occur when two coins are tossed. a. Derive the probability distribution of Y. -Mi. c. Derive the mean and variance of Y. . r \*± A \: .-, Vij

?

^ * \ * ° V

.»*

:

- ^ ,

c limate K mr=n d ££ f (y i . )T .^ ^ / '^ "Calculate E{Y\X =a 1)nand ^ i1^ ^= =o0 ^ T; - > ' v * * ' • » ' - ^

K V

b. Derive the cumulative probability distribution of Y. ., ,,
t

'v*2:

Use the probability distribution given in Table 2.2 to compute (a) E( Y) and E(X); (b) ax and ay\ and (c) tryy and corr(.\', Y).
dom

• y^1 d. Calculate the unemployment rate for (i) college graduates and (ii) non-college graduates. ^l"- v52 -^-.j.^ Q,w ^tl^\;^ to-lw, - _ o*^ e. A randomly selected member of this population reports being unemC\ % A . ployed. What is the probability that this worker is a college graduate? 33? A non-college graduate? ^ V \ \ - - ^ ^ \ ^ - ° ^ XG! _, t). ^^^A 1 ^-*^* 1 \ ^< \. " 1 f- Are educational achievement and employment status independent? explain
c v w

£(S \^. \\

^ " , - _ £ r - * t.^ _«•.*. ^\ 2 ® Using the random variables X and Y from Table 2.2, consider two new ran^ r J* - tCtf* ~ ^ - . < - ^ ^A
3 v

variables W ~ 3 + 6X and V -- 20 - 1Y. Compute (a) E(W) and E{V)\ (b) o-£ and o-^; and (c) tr^y and corr(W, V).

2.4
n

Suppose X is a Bernoulli random variable with P{X = 1) -= p.
a

^ t T ^ V ^ < ^ M

^ " i

-*

* Showi?(,V 3 )=p.

owE(A'*)=pforA; > 0. •« M * V« CT^ E [ ^ Pl-t" MIAU 1 j-i r j-' arf ^ . ^ o . ,A 9C c. Suppose that p - 0.3. Compute the mean, variance, skewness, and kurr a. [>-^v tosis of X. (Hint: You might find it helpful to use the formulas given in %v eWG* * oKiON1, - o Exercise 2.21.) ^ j ^ ^ 2.5 In September, Seattle's daily high temperature has a mean of 70°F and a o ^ = t D j ^-{oX^o-nr ( ^ to' n W ^ ' x ^ t a n d a r d deviation of 7°FWhat are the mean, standard deviation, and variance in °C? 2JS) The following table gives the joint probability distribution between employment status and college graduation among those either employed or looking for work (unemployed) in the working age U.S. population for 2008.

" *MHV i yw*<\ «w*vt6. -^o^A N^ ^^ -qs^* t^.^vs^ ^ ^v>o:fr^ ^ * ; ^ V A ^ ^ U V ^ . ? ^ x . ^^= V i t ^ x \ 7 M ~ " * cT-^ " ^ ^ ^ 2.7 In a given population or two-earner male/female couples, mare earnings N : -ie-^ ' ^ ^ have a mean of \$40,000 per year and a standard deviation of \$12,000. e ^VN Female earnings have a mean of \$45,000 per year and a standard deviation of \$18,000. The correlation between male and female earnings for a couple is 0.80. Let C denote the combined earnings for a randomly selected couple. a. What is the mean of C? b. What is the covariance between male and female earnings? c. What is the standard deviation of C? d. Convert the answers to (a) through (c) from U.S. dollars (\$) to euros ( !) . 2.8 2.9 The random variable Y has a mean of 1 and a variance of 4. Let Z = \{Y- 1). Show that \LZ - 0 and tr\ = 1. X and Y are discrete random variables with the following joint distribution:

58

CHAPTER 2 Review of Probability r 14 1 Value of X 5 8 0.02 0.17 0.02 22 0.05 0.15 0.03 Value of Y 30 0.10 0.05 0.*5 40 0.03 0.02 0.10 65 0.01 0.01 0.09 V ~\

x \ a- fa
;

Exercises

59

2.13 A is a Bernoulli random variable with P r ( A ^ 1) = 0.99, Y is distributed N(0,1), W is distributed N(0,100), and A, Y, and W are independent. Let lO\_ " 7 S - A T + ( l - A ) R ( T h a t i s , S - Y w h e n A = l , a n d £ = W w h e n A = 0.)
r

[ *51, ^ /"

^ ^ b

a. Show that £(Y 2 ) = 1 and £(W 2 ) = 100. b. Show that £( y 3 ) - 0 and £ ( IV3) - 0. (ffinc What is the skewness for a symmetric distribution?)

c Show that E( Y4) ~ 3 and E( W 4 ) = 3 X 1002. {Hint: Use the fact that \\ \ the kurtosis is 3 for a normal distribution.)
v r

T h a t i s , P r ( A = l , Y - 1 4 ) - 0.02, and so forth. a. Calculate the probability distribution, mean, and variance of Y. b. Calculate the probability distribution, mean, and variance of Y given A"-8. c. Calculate the covanance and correlation between A'and Y. M-\,cr £ 205^ Compute the following probabilities: 4TS ™ g ^ - l _ V v \ ' " p . If Yis distributed N(l,4),iind Pr(Y ^ 3). -\ \ W^l b. If Yisdistributed;V(3,9),findPr(y > 0). c. If Y is distributed /V(50.25),find Pr(40 < Y < 52). d. If Y is distributed N(5.2). find Pr(6 < Y < 8). 2.11 Compute the following probabilities: ^ 1 ^ ,/ 4 * > ^ 3 r p^-^p-V^ ~~ I 5 cf \2 N J \*' " i w T T a. If Y is distributed * i find Pr( Y < 7.78). \ b. If y is distributed ^ n . find P r ( K > 18.31). c. If y is distributed Fm^, find Pr( y > 1.83). d. Why are the answers to (b) and (c) the same? e. If Vis distributed* !, find P r ( y < 1.0). (Hint: Use the definition of the x\ distribution.) 2.12 Compute the following probabilities: r ^ s \c\ t-z — 12 a. If y is distributed r15, find Pr( Y > 1.75). b. If y is distributed f90,find Pr(-1.99 < Y < 1.99). c. If y is distributed N(0,1), find Pr(-1.99 < Y < 1.99) d. Why are the answers to (b) and (c) approximately the same? e. If Y is distributed F7.4, find Pr(Y > 4.12). f. If y is distributed F%m, find P r ( Y > 2.79).
2

W^-

r

<\T-

!>C "***

d

*

Derive

£ ( 5 ) - £ ( ^ 2 ) , £(S 3 ) and £ ( 5 4 ) . (Hint: Use the law of iterated expectations conditioning on A ^ 0 and A = 1.) ***? ^C-N^ fi-o^^ N

. *

AA

e. Derive the skewness and kurtosis for S.^ ^^ 4 i ^

-^ -HKJ

2.1^<In a population /xy = 100 and ay ~- 43. Use the central limit theorem t& answer the following questions: -" 1 a. In a random sample of size n = 100, find Pr( V < 101). b. In a random sample of size n = 165, find Pr( Y > 98). c. In a random sample of size n = 64, find Pr( 101 ^ Y < 103). 2CXS Suppose Y{, i -= 1 , 2 , . . . , n, are i.i.d. random variables, each distributed N(10,4).

-i

:

-i

V "U . ^

a. Compute Pr(9.6 , < J s 10.4) when (i) n =-- 20, (ii) n = 100, and b. Suppose c is a positive number. Show that Pr(10 - c < Y < 10 + c) . becomes close to 1.0 as n grows large. - -^* ^ ^ . ^*Y * * " 2 c. Use your answer in (b) to argue that Y converges in probability to 10. 2.16 Y is distributed W(5,100) and you want to calculate Pr( Y < 3.6). Unfortunately, you do not have your textbook and do not have access to a nor mal probability table like Appendix Table 1. However, you do have your computer and a computer program that can generate i.i.d. draws from the N(5,100) distribution. Explain how you can use your computer to compute an accurate approximation for Pr(Y < 3.6). 2.17 Yfri — l, .., «, are i.i.d. Bernoulli random variables with p - 0.4. Let Y denote the sample mean. a. Use the central limit to compute approximations for i. Pr( Y ^ 0.43) when n - 100. ii. Pr( Y < = 0.37) when n - 400

"t '

2, A ^

- , - ^ '. ^ \ ."V*\\

_

/^sD

60

CHAPTER 2 Review of Probability b. How large would n need to be to ensure that Pr(0.39 s Y < 0.41) > 0.95? (Use the central limit theorem to compute an approximate answer.) 2.18 In any year, the weather can inflict storm damage to a home. From year to year, the damage is random. Let Y denote the dollar value of damage in any given year. Suppose that in 95% of the years Y ^ \$0, but in 5% of the years Y - \$20,000. a. What are the mean and standard deviation of the damage in any year? b. Consider an "insurance pool" of 100 people whose homes are sufficiently dispersed so that, in any year, the damage to different homes can be viewed as independently distributed random variables. Let Y denote the average damage to these 100 homes in a year, (i) What is the expected value of the average damage Y? (ii) What is the probability that Yexceeds \$2000?

Exercises

61

2.22 Suppose you have some money to invest —for simplicity. \$1 -and you are planning to put a fraction w into a stock market mutual fund and the rest, 1 - w, into a bond mutual fund. Suppose that \$1 invested in a stock fund yields Rs after 1 year and that \$1 invested in a bond fund yields Rb, suppose that Rs is random with mean 0.08 (8%) and standard deviation 0.07, and suppose that Rb is random with mean 0.05 (5%) and standard devia tion 0.04. The correlation between Rs and Rb is 0.25. If you place a fraction w of your money in the stock fund and the rest, 1 — w, in the bond fund, then the return on your investment is R = wRs + (1 - w)Rb. a. Suppose that w -•= 0.5. Compute the mean and standard deviation of R. b. Suppose that w ~ 0.75. Compute the mean and standard deviation of R. c. What value of w makes the mean of R as large as possible? What is the standard deviation of R for this value of w? d. (Harder) What is the value of w that minimizes the standard deviation of Rl (Show using a graph, algebra, or calculus.) 2.23 This exercise provides an example of a pair of random variables A and Y for which the conditional mean of Ygiven Adependson AbutcorrfA, Y) = 0. Let X and Z be two independently distributed standard normal random variables, and let Y -= X2 + Z. a. S h o w t h a t Z f ( Y | A ) = A 2 b. Show that ixYLT 1. c Show thai E(AY) = 0. (Hint: Use the fact that the odd moments of a standard normal random variable are all zero.) d. Show that cov(A, Y) ~- 0 and thus corr(A, Y) -- 0. 2.24 Suppose Yl is distributed i-i.d. N(0, a2) for i - 1, 2 a. S h o w t h a t £ ( Y ? / o - 2 ) - l . b. Show that W - (l/o- 2 )^"-! Y2 is distributed Xlc. Show that E( W) = n. [Hint: Use your answer to (a).] d. Show that K - Y j n --1 is distributed t„ , n.

"V

2.19 Consider two random variables X and Y. Suppose that Y takes on k values y\,>..,yk and that A takes on / values x\ ,X[. a. Show that Pr( Y - y}) = S L i P r ( Y - y / | A ' ^ j : I - ) Pr(A = x,). [Hint: Use the definition of Pr( Y=-y } \X~ *,-).] b. Use your answer to (a) to verify Equation (2.19). c. Suppose that X and Y are independent. Show that <TXY = 0 and corr(A, Y) - 0. 2.20 Consider three random variables A, Y, and Z. Suppose that Y takes on k values yy,..., yk, that A takes on / values xh..., xh and that Z takes on m values zL>...,im. The joint probability distribution of A, Y, Z is Pr(A - x, Y - y,Z — z). and the conditional probability distribution of Y U given A and Z is Pr( Y - y \ X = x, Z = z) - -^T^x^^z^z) a. Explain how the marginal probability that Y = y can be calculated from the joint probability distribution. [Hint: This is a generalization of Equation (2.16).] b. Show that E( Y) = E[E(Y'\X, Z)\ [Hint: This is a generalization of Equations (2.19) and (2.20).] 2.21 A is a random variable with moments E(X), E(X2), E(X3), and so forth. a. ShowE(X~ ti)^E(X3)~3[£(X2)][E(X)] f 2[£(A)] 3 . 6[E(X)f[E(X2)]

b. Show E(X - fi)4 = E(X4) - 4[E(A)1[E(A 3 )] -f 3[E(X)f.

2.25 (Review of summation notation.) Let x-, .., xn denote a sequence of numbers,^,. . .,yn denote another sequence of numbers, and a, 6, and c denote three constants. Show that

Show that cov( Y^ Yj) -* pa} for /" # .49) /=] 2. Show that E(V2) > £(W 2 ). Let a = -<rXY/ vx and b = 1.27 A and Z are two jointly distributed random variables.Y) is a variance. Suppose you know the value of Z. To derive Equation (2.ixY)] + E[b2(Y ixY)2) = a 2 var(A) ^2a6cov(A.^ ) 2 ] = 6 2 ^ . Suppose that n = 2.tiy)} + /xxE(Y . [//mf:Let h(Z) = g(Z).29) follows from the definition of the expectation.. so from the final line of Equation (2.i±x)(Y.(Txy/al (2. Forrc > 2.Y]} = E{\b(X-fix)+c(V fxv)}[Y-nY}\ = E {[b(X ^ ) ] [ Y . Yn are random variables with a common mean p. and the fourth equality follows by the definition of the variance and covariance. This appendix derives the equations in Key Concept 2.1 D e r i v a t i o n o f Result s in K e y C o n c e p t 2.^ ) + 6(Y-/* y )] 2 } = E[a2(X .fxx)2} + 2E\ab(X. which (using the definition of the correlation) proves the correlation inequality.E ( A j Z ) .35). Rearranging this inequality yields 2.E(X\Z)}-h(Z). a.E{a + bX + cV)][Y. X HI «a ft2^-l-c22)'/+2aft2^+2flc2>'/+ i=i i=i HI n d. To derive Equation (2. When « is very large. F) + 62 var(K) = a2ax + 2abaxY + b2a^ (2. Applying Equation (2.33J.51) b. s o t h a t K = [ A .bam + ca-yy./Ay and var(Y) d. use the definition of the variance to write var(a + bY) £{[fl-^6Y-£(fl + 6Y)]2} = £{[6(Y-/* r )l 2 » = 6 2 f i [ ( Y .3..3 airy s (TX(Ty (covarianceinequality). X)' s 1.fxx) 4.l)fn]p<TY. (Hint: use the law of iterated expectations. 2.fiY) + fj.30). equivalent!^ \crXY/(frxa'Y)\ — 1. write E(Y2) = E{[(Y~fj.Y) + Ay]2} = E[{Y~py)2] + 2 i±YE{ Y. j corr (A.fj. where the second equality follows by collecting terms.[(« .34). use the definition of the covariance to write covffl + bX + cV. the third equality follows by expanding the quadratic.26 Suppose that Yu Y2. (2. y ( f l + ^ + c > .31). and \etW = X-X denote the error associated with this guess. that is. and V .51) it must be that aY .. To derive Equation (2.52) The covariance inequality implies that o-XY/(axaY) < 1 or.) ^^Y^IP^Yc. where / # . To derive Equation (2. Show that £ ( W Z ) = 0 . . (2.M y ] } + E {[c(VfLV)\[Y-pY]} . Equation (2.fix)(Y . b. c.py) + fiY=o-Y + fiy because E(Y.(anx-*-bfir)f} = E{[a(X . Derive E(V2). but not the value of A Let A = E( X\ Z) denote a guess of the value of X using the information on Z. | corr (A'Y)| < 1. Y) = E{[a + bX + cV.A A denote its error. it cannot be negative.x) + fj.YE(X. write E(XY) = E{[(X. and the same correlation p (so that the correlation between Yt and Y} is equal to p for all pairs i and /. Show that E(W)=Q.y. we have that var(aA +Y)= a2ax + try 4 2aaXY = (-VXYIVXY <rx + cr\$h 2(-<rXY/ax)<TXY = <Ty.showthat £ ( Y ) = p .m 62 CHAPTER 2 Review of Probability »• HI 4 t Derivation of Results in Key Concept 2.t±Y) =• 0.p.] < Because varfaA' 4. show that var( Y) » po-y.33).<rXY/a-x ^ 0. Let X = g(Z) denote another guess of A using Z./xxfXy = rrxy + (JLX I±Y. Show that £ ( Y ) .32).x][{Y fiy) + fiY]} = E[(X . a common variance oy. y a n d var(Y) = cr y /« J.3 To derive Equation (2. a.).I ) 2 = « f l 2 + /=i Ibc^XM var(oA + bY) = E{[(aX+bY) .31). .50) which is Equation (2. We now prove the correlation inequality in Equation (2. use the definition of the variance to write 63 ^axi^a^xt i=] (=i a HI J=I c.

Use the survey results to estimate p.i.0.5.-K -i \ .4 Using the data in Exercise 3. Use the central limit theorem to answer the following questions: i i . W . p > 0.1 Explain the difference between the sample average Y and the population mean. and the voters are asked to choose between candidate A and candidate B. significance level. b. . . Did the survey contain statistically significant evidence that the incumbent was ahead of the challenger at the time of the survey? Explain.-t . c In a random sample of size n — 165. 3. and (c) n = 1000. b. . Provide an example of each. What is the p-value for the test HQ: p = 0. b.v/\S 1 % & "H I C O ^ a.i. Let p denote the fraction of all likely voters who preferred the incumbent at the time of the survey. Let p denote the fraction of voters in the population who prefer candidate A. v\ 3. In a random sample of size\n =64JundPr( 101 < Y < 103).5? e.p(l standard error of your estimator.0. What is thep-value for the test 7 / 0 : p ^ 0. to calculate the fr-l Review the Concepts 3.5 A survey of 1055 registered voters is conducted.a? a. c. (b) -1.8 Explain why the differences-of-means estimator. H\ p * 0. Show that p is an unbiased estimator of p.50 at the 5% significance level. a. 3.0 *1 96 CHAPTER 3 Review of Statistics power of a test (77) one-sided alternative hypothesis (79) confidence set (79) confidence level (79) confidence interval (79) coverage probability (81) test for the difference between two means (81) causal effect (84) treatment effect (84) scatterplot(91) sample covariance (91) sample correlation coefficient (sample correlation) (92) t * V * z \ . Use the estimator of the variance o£p. Relate your answers to the law of large numbers. In a survey of 400 likely voters. What role does the central limit theorem play in statistical hypothesis testing? In the construction of confidence intervals? What is the difference between a null and alternative hypothesis? Among size. 5 ? d.i In a population. fiy — 100 and <rY = 43. H::p ^ 0 . Why do the results from (c) and (d) differ? f.6 Exercises (?.9. applied to data from a randomized controlled experiment.2 IA\ Ybe a Bernoulli random variable with success probability Pr( Y = 1) . 3. Show that p -= Y. f ' " * v n 4c' Exercises 97 . Sketch a hypothetical scatterplot for a sample of size 10 for two random variables with a population correlation of (a) 1.5 vs. L e t p be the fraction U of successes (Is) in this sample. (e) 0.4 3. (c) 0.5 vs. is io 3.p)/n. sample from this population for (a) n = 10.7 3. Construct a 99% confidence interval for p.K. (\$ w*> 4C\_ and let Y . Without doing any additional calculations. Hj. and let p denote the fraction of voters in the sample who prefer Candidate A. draws from this distribution. and power? Between a one-sided alternative hypothesis and a two-sided alternative hypothesis? Why does a confidence interval contain more information than the result of a single hypothesis test? 3. (b) n — 100. Show that var(p) = p ( l .50 vs. 215 responded that they would vote for the incumbent and 185 responded that they would vote for the challenger. . (d) 0. Yn be i. Construct a 95% confidence interval for p. find Pr(Y < 101).5 c.3: a. Why is the interval in (b) wider than the interval in (a)? d. 3. — p)/n. Determine the mean and variance of Y from an i.p. A population distribution has a mean of 10 and a variance of 16.3 Explain the difference between an estimator and an estimate.d.' 30. . \ WaA*»©S4A 3. rind Pr( Y > 98). i_ -* t' 5 X .2 3. test the hypothesis H0\ p = 0. x. is an estimator of the treatment effect.3 c.«?> ') b. .0. In a ranSoni samplritf size n^ 100. and let p be the fraction of survey respondents who preferred the incumbent.d.

Construct a 95% confidence interval for p. a. v n a given population. The sample average score Y on the test is 58 points. T w ^ .02. defined in Equation (3.d. producing a sample average o{62 feoints and o*. = S^versus / A ^ ^ j V s i n g the usual /-statistic yields a\p-value of 0.ffl* • ^ * * & W * -*wji -* t n e s t a n d a r d error of the difference in the two sample means? What .draws from a distribution with mean p. Construct a 50% confidence interval for p. i. What is the probability that the true value of p is contained in all 20 of these confidence intervals? ii. you wanted Pr(|p .p 1 > 0. using independently selected voters in each survey.l ^ T o inyestigate^ossible gender discrimination in a firm. c Suppose that the survey is carried out 20 times.^ sample standard deviation of 11 points.9 99 2eo &•-!* v>y\$as S 2© & . — 2000 vs. Yn be i. *\$»«*>£* ^ . What is the size of this test? ii. i.53. Construct a 95% confidence interval for the mean score of all New Jersey third graders. she will conclude that the new process is no better than the old process.5 vs. Test H0: p . is 8 points. sY. p. 11 % of the likely voters are African American. What is the power of the plant manager's testing procedure? ^cp« •.«X val? Explain. Suppose the new process is in fact better and has a mean bulb life of 2150 hours. iii. iv.54.^ * ' 'XlO4 Suppose a new standardized test is given to 100 randomly selected thirdgrade students in New Jersey. v. Compute the power of this test if p = 0. Suppose that you decide to reject H0 if \p — 0. /t i. ^fe-^pA ^ ^ ^ . a 95 % confidence interval for p is constructed. c^ \iv if - is the p-value of the test of no difference in means versus some difference?) 3. Let p denote the mean of the new process. A summary of the resulting monthly salaries follows: >^Q t Vol . How large should n be if the survey uses simple random sampling? Let Yj. S .01) < 0.25o-y/«. p = 0. That is. Does the 95% confidence interval contain p.2 ^ ^ ? ? ^ J b. p. H\: p > 0. — 6is contained in the 95% confidence inter. You are interested in the competing hypotheses H0: p = 0. A surey using a simple random sample of 600 landline telephone numbers finds 8% African Americans.03..p. ii.96 X SE(p). For each of these 20 surveys.i. v\~ \. She says that she will believe the inventor's claim if the sample mean life of the bulbs is greater than 2100 hours. . a sample oy. Construct a 99% confidence interval for p. a. Show that (a) E[Y) .. The plant manager randomly select^ 100 bulbs\produced by the process.* £ „ *l* JJgraders from Iowa. In the survey.co iro . _ =.Y and (b) var(Y) = 1.11 Consider the estimator Y. Construct a 90% confidence x interval for the difference in mean scores between lov^a and New Jersey. p = £ 0. A / i A.5.*P ^ tp* Suppose that a lightbulb manufacturing plant produces bulbs with a mean life of 2000 hours and a standard deviation of 200 hours. What is the size of the plant manager's testing procedure? w 2^ \ •KJ P V V - b. The authors plan to administer the test to all third-grade students in New Jersey. otherwise.5 using a 5% significance level.5 [ > 0. a.^ Qan v o u c o n c i u d e with a high degree of confidence that the popula* 1 L A C ( V ' \<?h 4vV* V^. b. In survey jargon. that is. The sample mean test score is 1110. Consider the null and alternative hypothesis H0: p. i. Hy. An inventor claims to have developed an improved process that produces bulbs with a longer mean life and the same standard deviation. = 5? Explain. Hj.1). > 2000. 3\$ A new version of the SAT test is given to^lp00o.Suppose the same test is given tq200_|andomly selected third V ~Z . What testing procedure should the plant manager use if she wants the size of her test to be 5%? ***" J> **"V* *~* .98 CHAPTER 3 Review of Statistics a.5 vs. it is half the length of 95% confidence interval. //i: p ^ 0. Construct a 95% confidence interval for the population mean test score for high school seniors. Test H0: p — 0. r Exerdses \ 3.5 using a 5% significance level.5 vs.0. Suppose you wanted to design a survey that had a margin of error of at most 1 %. Can you you determine ii p. Is there evidence that the survey is biased? Explain. A test of HQ. b.00)men and|64wgnieji with similar job descriptions are selected at random.* A ^ < S ottl .^ tion means for Iowa and New Jersey students are different? (What is . How many of these confidence intervals do you expect to contain the true value of p? d. the "margin of error" is 1. \&& .05. and the sample standard deviation. and the sample standard deviation is 123.andomly selected high school seniors.

14. Let population a denote the "opt-out" (treatment) group and population b denote the "opt-in" (control) group. ^ i. Remember that the samples are independent. Construct a 95% confidence interval for the change in average X' V ^ *5> test score associated with the prep course.73 in. a.5 in. Construct a 95% confidence interval for the treatment effect. Is there statistically significant evidence that Florida students perform differently than other students in the United States? V o* y * * ' '" c.15 Let Ya and Yb denote Bernoulli random variables from two different populations.. and finally.p 6 ) = P a ( 1 " Pt<) \$ . Show that var(p fl .9 n 238 182 fc + £>£L^A (/finr. The average change in their test scores is 9 points.100 CHAPTER 3 Review of Statistics Average Salary ( V) Men Women \$3100 \$2900 Standard Deviation (sY) \$200 \$320 n 100 64 ^ Exercises 101 ^M. The test is administered to 453 randomly /Us1 selected students in Florida.0 Standard Deviation (sY) 19.) c Suppose that naand nb are large.y .16 Grades on a standardized test are known to have a mean of 1000 for students in the United States. Suppose the sample from population a is independent of the sample from population b. X lb.^vTCcA YL 3CC X. Suppose that E(Ya) — pa and E( Yb) = pb. second. Construct a 95% confidence interval for the average test score for Florida students. Y .4 650.158 lb.2 and standard deviation sY .4 17. They are v . sY .. compute the p-value associated with the r-statistic. with sample average denoted pa.r st>_ Average Score {Y) 657. What do these data suggest about wage differences in the firm? Do v they represent statistically significant evidence that average wages of k>^ men and women are different? (To answer this question. pa . b.. sx = 1. Construct a 95% confidence interval for the change in average test scores. ^ a. -3 i i. and a random sample of size nb is chosen from population b. first state the null and alternative hypothesis. Show that E{pb) -^pb andvar(pb) = ph(l -pb)/nh. i. Read the box "A Novel Way to Boost Retirement Savings" in Section 3.pa)/na. Show that a 95% confidence interval Pb forp fl -pb is given by (pa .i\l~ 3. Do these data suggest that the firm is guilty of gender discrimination in its compensation policies? Explain. Construct a 95% confidence interval for the mean test score in the population.2W30 r ^~ 2o> Sv4= ?ao > **^-M„t-Q '5£S>^/ ^3 ^\z*%- with sample average denoted pb. 3. themeanis 1013 and the stan™s\& dard deviation (s) is 108. a. * • > ^ r\ A * v ** 0 ^ ^ 3 M % l- VJ> a. use the p-value to answer the question."V-^^ districts in California yield Y = 646.96. <s^£* * % 1 .8 in. & K given a 3-hour preparation course before the test is administered.2 lb.pb. Another 503 students are selected at random from Florida.1 9 .. Their average test score is 1019 with a standard deviation of 95.13 Data on fifth-gr'ade test scores (reading and mathematics) for SChOOl a. ^ ^ .pb) ± 1.14 Values of height in inches (Xs) and weight in pounds (Y) are recorded from a sample of 300 male college students.5. 3. and the standard deviation of the change is 60 0 points. s. Convert these statistics to the metric system (meters and kilograms) . When the districts were divided into districts with small classes ( < 20 students per teacher) and large classes ( ^ 20 students per teacher). b. third. '. . Is there statistically significant evidence that the prep course helped? d The original 453 students are given the prep course and then are d.) b. denoted a and b. 3./ P B ( 1 ~ ^ + ^~Pb\ n l V a 'b How would you construct a 90% confidence interval for pa — pbl d.WX. The resulting summary statistics are X = 70. r ig^i&3. A ran dom sample of size na is chosen from population a. yVj-v «« »™.0-85. and r XY .in this sample. the following results were found: Class Size Small L>arge v.Afc ^ <e\ ^ to* \^i < \ V £ i ii. asked to take the test a second time. compute the relevant f-statistic. b. Sxy = 21. Show that E(pa) ~-pa and var(p tt ) -~pa(\ . 5 ^ 4 ^ . Is there statistically significant evidence that the districts with smaller classes have higher average test scores? Explain.

/ Exercises y 4.. ' U J J . and between the OLS predicted value If and E(Yt\Xi). There are three key assumptions for the linear regression model: (1) The regression errors. and this year it has 23 students. SER = 11.5 in. xrI k \« ^ \ l ' _i . What is the regression's predict! qn_for. T V a population and that these men's height and weight are recorded. What is thejegression's prediction for that classroom's average test score? -Vn^Suf = _3*-«3 ^ . ^ i '*•* r -ft.. (2) consistent.5. Last year a classroom had 19 students.•&** •F**• t? A-^ . A classroom has 22 students.\ A c. 3. What is the sample average of the test scores across the 100 classrooms? (Hint: Review the formulas for the OLS estimators.d.The OLS estimators of the regression intercept and slope are denoted fa and fa..t.»^v ~°£\ \ AN ! * ^ ^ ' • • ---". have a mean of zero conditional on the regressors X. The population regression line can be estimated using sample observations (Yh Xt). _ M OLS regression line (114) sample regression line (114) sample regression function (114) predicted value (114) residual (115) regression^? 2 (119) explained sum of squares (ESS) (119) total sum of squares (TSS) (119) sum of squared residuals (SSR) (120) standard error of the regression (SER) (120) least squares assumptions (122) 2. estimated coefficients. SER •= 10.520.81..% 1 \~JW "^ * Key T e r m s linear regression model with a single regressor (110) dependent variable (110) independent variable (110) regressor (110) population regression line (110) population regression function (110) population intercept (110) population slope (110) population coefficients (110) parameters (110) error term (110) ordinary least squares (OLS) estimators (114) Review the Concepts •* 4.A*iS*^ ^yg a i a t e growth spurt and grows 1.2 Explain the difference between fa and fa. yjtf \5 a. over the course of a year. random draws from the population. f TestScore .82 X CS.5.-99.1 Suppose that a researcher.i-iviiugiuui regression? (Give all results.= t*y sion of weight on height yields | j— Y*£>> l ^ Weight . i .2. s. Suppose that instead of measuring weight and height in pounds and inches these variables are measured in centimeters and kilograms. a.5.the . tall? 74 in.08. estimates the OLS regression Jk. The sample average class size across the 100 classrooms is 21. using data on class size (C5) and average test scores from 100 third-grade classes. ut..1. tall? 65 in. What is the sample standard deviation of test scores across the L O O classrooms? (Hint: Review the formulas for the R2 and SER. If these assumptions hold. A ^ ^ ^ . St£{>-"Td.s V i * i b. The R2 and standard error of the regression (SER) are measures of how close the values of Y.-where Weight is measured in pounds and Height is measured in inches.) S £ & * ..change in the classroom average test score? A • -*5 * .J 4.> ^ * o V S ) ^ .. between the residual ut and the regression error «.i. n by ordinary least squares (OLS). 4. the OLS estimators fa and fa are (1) unbiased.9. then provide an example in which the assumption fails. The standard error of the regression is an estimator of the standard deviation of the regression error._ ^ man .94 x Height. Sketch a hypothetical scatterplot of data for a regression with R2 = 0. with a larger value indicating that the Y/s are closer to the line.4.) —•^ Suppose that a random sample of 200twenty-year-old men is selected from • a . proyide an example in which the assumption is valid. tall? (.41 +." What are the regression estimates from this new ceptimeter-kilogram mmius nuiii una new L. and (3) large outliers are unlikely.J .. ' % s-oxxw \*t.3 Exercises 133 Sketch a hypothetical scatterplot of data for an estimated regression with R2 = 0.3. WW. 0.^^56. A regres£..cjrtJ*uciE. and (3) normally distributed when the sample is large. Ryand SER.. are to the estimated regression line.4 .) . For each least squares assumption.% (2) the sample observations are i. 4. The R2 is between 0 and 1. R2 = 0. R > 2 -. ^ ^ate What is the regression's prediction for the increase in this man's weight? c.132 CHAPTER A Linear Regression with One Regressor 2.What is the regression's weight prediction for someone who is 70 in.

) c.2.4 W£ = 696. and is it consistent \ . .4 Read the box "The 'Beta' of a Stock" in Section 4. 5 W 6 > < g k V*. ' l^c^V&fcvtf^^^V^ ^rfwfc ^ o r 6 S ote w & f c ^ G**^ c# ^ r e tthe Are other assumptions assumptions in in Key Key Concept Concept 4. Derive a formula for the least squares estimator of j3>. . a. Derive an expression for the large-sample variance of j81.Rf) for this stock is greater than the variance of ( R m . For each company listed in the table in the box. Explain what the coefficient values 696. which is shown in Appendix 4.3%. where (A). Does this imply that fa . and consider the regression model Yt = fa + faX{ + uh ^ Explain what the term «. When X ^ 1. Compute the estimated gain in score for a student who is given an . (Hint: Use the fact that £ : is unbiased. ^ ^ ^ S O n ^ ii.0. Suppose that the value of {3 is greater than 1 for a particular stock.).measured in pollarsjon p. u{) are i. a.^ "^v y ™ & 3 * ^ i cipSCx ?* f. What are the units of measurement for the R21 (Dollars? Years? Or is R2 unit-free?) d.24 Xj. -". one ^ ot the examination times based on the flip of a coin. 4. a. the rate of return on 3-month Treasury bills is 3. and X{ is a 7 7 ^ . -— 1 ^ 4. .9.^ .4 continue to hold? Which change? Why? (Is £j normally distributed in large samples with mean and variance given in Key Concept 4. * d s^*v> y (biOto^ Q«>C^ei**A u 4 A . \ i .] 4.3 4. Show that R~ . use the estimated value of p to estimate the stock's expected rate of return.) 4.KstA<£ v W M ^ . Suppose that the value of /? is less than 1 for a particular stock. i. The regression R2 is 0. . . implies that ft^^Ztt^jji^ plXh = f}Q + 4.6 years.^ X Age. Let Yt denote the ~j • number of points scored on the exam by the Ith student (0 s Yt ^ 100). .R.d. Compute the estimated regression's prediction for the average score of students given 90 minutes to complete the exam. CHAPTER A Linear Regression with One Regressor 1% .ut..)? (Hint: Don't forget the regression error. SER = 624. The standard error of the regression (SER) is 624.20. A linear regression yields fi\ s 0.7 and 9.3 A regression of average weekly earnings (AWE. A linear regression yields R2 .21). Show that the regression assumptions in Key Concept 4.***i'*Je*-r ^ ^ x & J .-. L\\ t e ^ ^ ^ S £MV^ w ^ o \ s Exercises 135 .i.6 Show that the first least squares assumption.9 ( £H^ Suppose that Yj = fa + faXt t-«.0.. ^ ^ V e t v x s * Will the regression give reliable predictions for a 99-year-old worker? _ Wh ^ * ^ • <*\ ^ tsuAV^ * * V or why not? tf: O ^ ^ g g ^ ^ .7 4. b. . -*. w. E(ui Xt) .4). 4. but some students have 90 minutes to complete the exam while others have 120 minutes.49 + 0..2. Suppose you know that fa — 0. satisfied? Explain.represents. b. b.5% and the rate of return on a large diversified portfolio of stocks (the S&P 500) is 7.[i"gi~ . [Hint: Evaluate/the terms in Equation (4.8 Show that fa is an unbiased estimator of fa. What are the units of measurement for the SER1 (Dollars? Years? Or is SER unit-free?) c. a. He gives each of the 400 studentsjn his course the same final exam.11 Consider the regression model Yj = fa + p^X.1.023.023. let Xt denote the amount of time that the student has to complete the exam {X. n e other d.) Suppose that all of the regression assumptions in Key Concept 4. Show that the variance of (R . additional 10 minutes on the exam.< S" & ^ age (measured in yearsj_using a random sample of college-educated full&<\$ * ^ * \ ^ towOA^^ V j ^ t t i m e w o r k e r s a g e d 25-65 yields the following: SE*» V ^^ ' ^ ^ Of « * ^ f ^ e _ _ " '* K S£fc . ^ (Bernoulli random variable with Vr(X = 1) = 0.4? What about £ 0 ?) a.0? & v.7 1. The estimated regression is >. Given what you know about thetjt^Kbution of eanungg^o you think 0«*M V^ ^ f *^ \$ & * ^ ^ ijrt&f\ it is plausible that the distribution of errofsMfftne'regression is norc a ^ J o V cio._ skewed? What is the smallest value of earnings. z _^ JA >' .-. Repeat for 120 minutes and 150 minutes./?.3 are satisfied except that the first assumption is replaced with E(ut Xt) — 2.5 A professor decides to run an experiment to measure the effect of time pressure on final exam scores.0. Which parts of Key Concept 4. What is the average value <&<S&& oc AUKS of A WE in the sample? (Hint: Review Key Concept 4.AO* with a normal distribution?) ^ .G?tfo n ~ . Wl A t £* '" ^ v i B ^ 11 fa A S = \ C T J ] A 45-year-old worker? e W° Sa-^s^j ^ ^ e .«-^\axi wJ^cw«i (\\>J^)p^-*aV<>irt 8* The average age in this sample is 41. Why will different students have different values of u{l Explain why E(ui\Xt) = 0 for this regression model. R2 . ^L* g_„.J ^ .3.134 . = 90 or 120). is N(0.3 are satisfied.6 mean.0. Is it possible that variance of (R Rf) for this stock is greater than the variance of (Rm . What is the regression's predicted earnings for a 25-year-old worker? «•* ~ V ^ f c n * ^ ! <=*o^ . b. In a given year.(e —H ° ^ . Each student is randomly assigned -% * o 4. K ^ r mal? (Hint: Do you think that the distribution is symmetric or r . J ^ v^ w ^ Q<v.^ y * S * ^ &iN! v*ftv**A o ^ o W .1.

A. age 25-34.13 Suppose that 1.pearsonhighered./B.21) multiplied by jr.136 CHAPTER A Linear Regression with One Regressor b.) a. (Pr6ximity to college lowers the cost of education. What is the estimated intercept? What is the estimated slope? Use the estimated regression to answer this question: How much do earnings increase as workers age by 1 year? b. + KUi} where K is a non-zero constant and (Yh X{) satisfy the three least squares assumptions. (Hint: What is the sample mean of Beauty!) c.1 On the text Web site http://www. as their highest degree. also available on the Web site. .S. Run a regression of average hourly earnings (AHE) on age (Age).Y). Show that the regression R2 in the regression of Y on X is the squared value of the sample correlation between X and Y. where rXY is the sample correlation between X and Y. Predict Professor Stock's and Professor Watson's course evaluations. Suppose you know that fa = 4. a. E4.com/stock_watson/. across courses? Explain.com/stock_watson/. August 2005.' = fa + faX. on average. "Beauty in the Classroom: Instructors' Pulchritude and Putative Pedagogical Productivity. (Generally. course characteristics. Comment on the size of the regression's slope. That is. Does there appear to be a relationship between the variables? b. It contains data for full-time. 1 ~?r ^— * • . var[(X.com/stock_watson/. In this exercise./0". Does Beauty explain a large fraction of the variance in evaluations . -. Bob is a 26-year-old worker Predict Bob's earnings using the estimated regression.14 Show that the sample regression line passes through the point (X. you will find a data file TeachingRatings that contains data on course evaluations." e. Alexis is a 30-year-old worker. full-year workers. 4. variance of /3i is given by crx — K « • " ^ -. Does age account for a large fraction of the variance in earnings across individuals? Explain.24(4): 369 -376. (These are the same data as in CPS92_08 but are limited to the year 2008. complete These data were provided by Professor Daniel Hamermesh of the University of Texas at Austin and were used in his paper with Amy Parker. One of the characteristics is an index of the professor's "beauty" as rated by a panel of six judges. older workers have more job experience. Construct a scatterplot of average course evaluations (Course_Eval) on the professor's beauty (Beauty). and sY and sx are the sample standard deviations of X and Y.3 On tne text Web site http://www. In this exercise.[Hint: This equation is the variance given in equation (4. leading to higher productivity and earnings.12 a. c.] 4. Predict Alexis's earnings using the estimated regression.2 Empirical Exercises 137 On the text Web site http://www.1 for 2008. Professor Watson has an average value of Beauty. so that students who-live closer to a four-year college should. while Professor Stock's value of Beauty is one standard deviation above the average. you will find a data file CPS08 that contains an extended version of the data set used in Table 3. you will use these data to investigate the relationship between the number of completed years of education for young adults and the distance from each student's high school to the nearest four-year college. Derive a formula for the least squares estimator of Bv 4. K4.) In this exercise. Show that the large sample » .pearsonhighered. you will find a data file CollegeDistance that contains data from a random sample of high school seniors interviewed in 1980 and re-interviewed in 1986. Run a regression of average course evaluations (Course _Eval) on the professor's beauty (Beauty). Is the estimated effect of Beauty on Course_Eval large or small? Explain what you mean by "large" and "small. you will investigate how course evaluations are related to the professor's beauty." Economics of Education Review. with a high school diploma or B. b. and professor characteristics for 463 courses at the University of Texas at Austin. Empirical Exercises E4. show that R2-r2XY. d. you will investigate the relationship between a worker's age and earnings. What is the estimated intercept? What is the estimated slope? Explain why the estimated intercept is equal to the sample mean of Course_Eval.pearsonhighered. A detailed description is given in CPS08_Description. Show that the R2 from the regression of Y on X is the same as the R2 from the regression of X on Y c. also available on the Web site. 1 A detailed description is given in TeachingRatings ^Description.""Ba^ . Show that fa = rXy(sY/sx).

2. a variable that is equal to 1 if the person is female and 0 if the person a inale. . Define the wage gender gap as the difference in mean earnings between men and women. using wage data on 250 randomly selected male o^>.1 . Do you reject the null hypotKesis at the_5% level? At the 1% level? \ . SER = 4. Another researcher uses these same data but regresses Wages on Female. 3° workers and 280 female workers. Construct a 99% confidence interval for fa.1 Suppose that a researcher.4 5. Provide a hypothetical empirical example in which you think the errors would be heteroskedastic and explain your reasoning. .t *¥> * Cy -£ -5Z -. ^ H0: fa = -5. R2 ~. R2 = .52 + 2.0.1 Outline the procedures for computing the p-value of a two-sided test of H0: p. .\A i<A n Y l -ft { £ -o ^•S<o where Wage is measured in dollars per hour and Male is a binary variable that is equal to 1 if the person is a male and 0 if the person is a female.81. . ti2 = 0. SER = 10.^2. what is the mean wage of women? Of men? \l . R2 -.6.. estimates the OLS regression X / Wage = 12. set of observations (Yh X. estimates the OLS regression TestScore = 520. Explain how you could use a regression model to estimate the wage gender gap using the data on earnings of men and women.).~S & population and their heights and weights arerecorded.12 X Male.0.21) { ^h? - 7< o * 3 6^ ?? 4~ 5. What is the estimated gender gap? i \L X '.^ . v 'Xj Suppose that a researcher.i. (2. n. Construct a 95% confidence interval for the gender gap.6 is contained in the 95% confidence interval for Bv.d.06.-?-\j. i\~ 1 . .82 X CS. SER . yd. n.23) (0.2 5. a.S~k } \*\ e.5.41 f 3. (20.i._ V « f \ :&> b. Construct a 9^% confidence interval for the person's weight gain. A man has a late growth spurt and grows 1. l % rfj ^ b.36) Review the Concepts 5.y = 0 using an i.-J-2 AT. Calculate the p-value for the two-sided test of the null hypothesis \. A regression of weight on height yields oS /' V Os» 7 5* 3*v<o-y\ ^ / Weight = -99.X Female. set of observations Yh i = 1 .168 CHAPTER 5 Regression with a Single Regressor: Hypothesis Tests and Confidence Intervals coefficient multiplying Dt (154) coefficient on D-t (154) heteroskedasticity and homoskedasticity (156) homoskedasticity-only standard errors (158) heteroskedasticity-robust standard error(159) Gauss-Markov theorem (162) best linear unbiased estimator (BLUE) (163) weighted least squares (163) homoskedastic normal regression assumptions (164) Gauss--Markov conditions (176) \ ' " A -• Exercises 169 c. \ ^ c * .3 Exercises 5. M -1& 'V- a.31) \ .~ty^ •** whether -5.2. Without doing any additional calculations. Outline the procedures for computing the p-value of a two-sided test of HQ: fa = 0 in a regression model using an i. using data on class size (CS) and average test scores from 100 third-grade classes. determine . W . . What are the dependent and independent variables? Define homoskedasticity and heteroskedasticity. What are the regression estimates calculated from this regression? * Wage =--£. fit-***'* .H ^ J t ^ r where Weight is measured in pounds and Height is measured in inches. . In the sample.94 X Height. the regression slope coefficient.11. Calculate the p-value for the two-sided test of the null hypothesis H0: & = 0. s t ^ * ? .o©r <^y> W<. Construct a 95% confidence interval for fa.3-*\Suppose that a random sample of 200 twenty-year-old men is selected from . SER ~ f. (.d.4) (2. Is the estimated gender gap significantly different from zero? (Compute lhe/?-value for testmg the null hypothesis that there is no gender c. '-A •\$ 5. d.5 inches over the course of a year.15) (0.08.

Show that 0 is conditionally unbiased.7 ) Suppose that (Y„ Xf) satisfy the assumptions in Key Concept 4.3 and. b. A sample of size n . and small classes contained approximately 15 students. A random sample of size n = 250 is drawn and yields ^J Y .4. R.2..170 CHAPTER 5 Regression with a Single Regressor Hypothesis Tests and Confidence intervals 5. Do small classes improve test scores? By how much? Is the effect large? Explain.5Xi In the 1980s. Hx: fa > 55 at the 5% level. ^ ^ ^ ^ p r ^ v VcjVW ^ ^ . ^ .6" \ Refer to the regression described in Exercise 5. Let SmallClass denote a binary variable equal to 1 if the student is assigned to a small class and equal to 0 otherwise.54. <y> where u{ and X{ satisfy the assumptions in Key Concept 4.4 Read the box "The Economic Value of a Year of Education: Homoskedasticity or Heteroskedasticity?" in Section 5. and given standardized tests at the end of the year. R2 = 0.. (10. Test H0: fa = 0 vs.2X. Is the estimated effect of class size on test scores statistically significant? Carry out a test at the 5% level.26. SER (1.30 yields ^ Y = 43. 5.5) 74.4 f 3.A ^ W ^ 0 ^ \$ S .5JT. Y„.2) (7.52. r « t u \ ^ \ 1 U A i 4 l c. How much is this worker's average hourly earnings expected to increase? c. fa * 55 at the 5% l e v e l .5) a. a.2 + 61. A randomly selected 30-year-old worker reports an education level of 16 years. c Suppose you learned that Yj and X{ were independent. a. (3. in the population. a.9 Consider the regression model vj 5t 5.YjA\ where Y and X are the sample means of Yt and X„ respectively. Suppose that Yi and Xi are independent and many samples of size n = 250 are drawn. SE(fa) was computed using Equation (5. SER = 1. In what fraction of the samples would H0 from . in addition. (Regular classes contained approximately 24 students. ^ " ^ 5. college grad uates earn \$10 per hour more than high school graduates. Do you think that the regression errors plausibly are homoskedastic? Explain.1) (1.3.0 0 1 . a. Let ~\$ denote an estimator of /? that is constructed as 0 . Show that 0 is a linear function of Yh Y2). o ^ y b.3).3. R2 = 0. What is the worker's expected average hourly earnings? b. A regression of TestScore on SmallClass yields TestScore .918. and (a) and (b) answered. Would you be surprised? Explain. 5^© •*&. a. H< fa^Qat the 5% level. on average..9 X SmallClass.2 ^ ° ^ ^ ° w >«l£ Su*»-2> KW>' ^ rt \ \5^ f 2 ^ ( : %s* b. SER -= 6. d.5.. the standardized tests have a mean score of 925 points and a standard deviation of 75 points. ^ ^ c.23) to answer the following. Construct a 99% confidence interval for the effect of SmallClass on test score. Test H0: fa = 55 vs. Construct a 95% confidence interval for fa. u( is N(0. Use the regression reported in Equation (5. H^..55 vs. Is this statement consistent with the regression evidence? What range of values is consistent with the regression evidence? r Exercises 171 5.6) (2.5. regressions estimated.4) ^wKeYelhe numbers in parentheses are the homoskedastic-onlv standard errors fof the regression coefficients. .0 + 13. b. Suppose that the regression errors were homoskedastic: Would this affect the Validity of the confidence interval constructed in Exercise 5. a2) and is independent of X.6. Test //„: fa . A high school graduate (12 years of education) is contemplating going to a community college for a 2-year degree. A high school counselor tells a student that.(a) be rejected? In what fraction of samples would the value B: — 0 be included in the confidence interval from (b)? Suppose that (Yh %) satisfy the assumptions in Key Concept 4. Tennessee conducted an experiment in which kindergarten students were randomly assigned to "regular" and "small" classes. y . b.5(c)? Explain.) Suppose that. Construct a 95% confidence interval for fa.

suppose that Y} denotes earnings.1.. where («.172 CHAPTER 5 Regression with a Single Regressor: Hypothesis Tests and Confidence Intervals Empirical Exercises 5. Showthatjhe standard error of (3mi . Is the estimated regression slope coefficient statistically significant? That is. 5%.10. Let Women denote an indicator variable that is equal to 1 for women and 0 for men and suppose that alf 251_^bservations are used in the regression Yt — fa 4 faWomen.o + BwlXwi + uwj. Repeat (a) using only the data for college graduates.) E5. fa + fa-. Yn. .3? 5.-. Repeat (a) using only the data for high school graduates.% 5.= x) is constant? d. B„ti denote the OLS estimator constructed from the sample of women.31). Construct a 95% confidence interval for the slope coefficient. c.fatl is given by S E ( / W ~k.10 and sw ..1 Using the data set CPS08 described in Empirical Exercise E4..15 A researcher has two independent samples of observations on (Yb X^)..\$51. c. 5%.i) = V[SE(pmtl)V + [SE(fa7l)i2. run a regression of average hourly earnings (A HE) on Age and carry out the following exercises.28) in Appendix 5. run a regression of Course_Eval on Beauty. or 1% significance level? What is the p-value associated with coefficient's /-statistic? b. (Hint: See Exercise 5. Xj) satisfy the Gauss-Markov conditions given in Equation (5. cr2u) and is independent of Xj..2 5.14 Suppose that Yj = BXt -f «. Derive the conditional variance of the estimator.1.3.-. How would your answers to (a) and (b) change if you assumed only that (Yn X^ satisfied the assumptions in Key Concept 4. The fcample averagejof men's weekly earnings [Ym^ (l/n m )2w=i^.15.3 and.% and fa = YX. Write the regression for men as Ymi . How would your answers to (a) and (b) change if you assumed only that (Yh Xi) satisfied the assumptions in Key Concept 4. Empirical Exercises E5. a. f ut.12 Starting from Equation (4. Is the effect of distance on completed years of education different for men than for women? (Hint: See Exercise 5.Bm0 + BmlXmi + umi and the regression for women as Kj ~ Pw.1.3 and var(«. Construct a 95% confidence interval for the slope coefficient. Show that fa = Y0.2 Using the data set TeachingRatings described in Empirical Exercise E4. can you reject the null hypothesis H0: /3I = 0 versus a two-sided alternative at the 10%. Is the estimated regression slope coefficient statistically significant? That is. d. e. can you reject the null hypothesis H0: B[ — 0 versus a twosided alternative at the 10%. Find the OLS estimates of 8\$ and Bx and their corresponding standard errors. or 1% significance level? What is the p-value associated with coefficient's r-statistic? b. c. 5.15.2. a ^ t h e ^ ^ g s f j n ^ a i r l j h ^ t j ^ i [sm = V^^I^(YmJ . Derive the least squares estimator of (5 and show that it is a linear function of Y±. can you reject the null hypothesis H0: B-^-Q versus a twosided alternative at the 10%.) E5.11 A rapdoin^amrjle of workers contains nm = 120 men and nw ~ 131 women. Prove that the estimator is BLUE. Show that the estimator is conditionally unbiased. The corresponding values for women are Yw ~ \$485. Xt denotes years of schooling.-] is \$ 5 2 3 . or 1% significance level? What is the p-value associated with coefficient's r-statistic? Using the data set CoIlegeDistance described in Empirical Exercise E4. and SE(j3ml) and SE(fatl) denote the corresponding standard errors. d. derive the variance of fa under homoskedasticity given in Equation (5.Ym)2] is \$68. 1 0 . d. a..22). Let if) denote the sample mean for observations with X = 0 and yj denote the sample mean for observations with X~\.«. e. a. Is the effect of age on earnings different for high school graduates than for college graduates? Explain. Run the regression using data only on females and repeat (b). 5. Is B i conditionally unbiased? b. 5%. Is the estimated regression slope coefficient statistically significant? That is.j^. run a regression of years of completed education (ED) on distance to the nearest college (Dist) and carry out the following exercises. in addition.13 Suppose that (Yh Xt) satisfy the assumptions in Key Concept 4. Let jSMil denote the OLS estimator constructed y . a. Is fa the best linear conditionally unbiased estimator of B{1 c. To be specific. ^_ 173 V " kPS ?•* ^ & & A 9~+ using the sample of men.10 Let X-t denote a binary variable and consider the regression Yt~ fa + faXi +. and the independent samples are for men and women. Run the regression using data only on males and repeat (b). b. ut is N(Q.

holding X2 constant. which occurs when one regressor is an exact linear function of the other regressors. The worker's ages ranged from 25 to 34 years. Give two examples of a pair of perfectly multicollinear regressors.2 Exercises 205 A multiple regression includes two regressors: Y{ = fa -f faX\i + faX2i + «. What is the expected change in Y if X\ increases by 3 units and X2 is unchanged? What is the expected change in Yif X2 decreases by 5 units and Xj is unchanged? What is the expected change in Y if Xy increases by 3 units and X2 decreases by 5 units? Explain why two perfectly multicollinear regressors cannot be included in a linear multiple regression. 0 otherwise) West ~ binary variable (1 if Region .The data set consists of information on^OOOfilfTtime full-year workers. 6. 4.. Do workers with college degrees earn more. Using school district data like that used in this'6hapter. The other regression coefficients have an analogous interpretation.4 Key T e r m s omitted variable bias (180) multiple regression model (186) population regression line (186) population regression function (186) intercept (186) slope coefficient of Xu (186) coefficient on Xu (186) slope coefficient of X2j (186) coefficient on X2i (186) holding X2 constant (187) controlling for X7 (187) partial effect (187) population multiple regression model (188) constant regressor (188) constant term (188) homoskedastic (188) heteroskedastic (188) ordinary least squares (OLS) estimators of fa. Do men earn more than women on average? How much more? 6. the R2. is it biased up or down? Why? 6. consistent. Perfect multicollinearity.204 CHAPTER 6 Linear Regression with Multiple Regressors associated with a 1-unit change in X*. Bh. mar ital status. 0 otherwise) 6.3 Using the regression results in column (2): a. than workers with only high school degrees? How much more? °^ ctvD* ^ >w-r^ b. 6. a. on average. 3. Solving perfect multicollinearity requires changing the set of regressors. usually arises from a mistake in choosing which regressors to include in a multiple regression. The data set also contained information on the region of the country where the person lived.West. The standard error of the regression.binary variable (1 if Region = Northeast. let AHE = average hourly earnings (in 1998 dollars) College . holding the other regressors constant.1 A researcher is interested in the effect on test scores of computer usage.Bk (190) OLS regression line (190) predicted value (190) OLS residual (190) R2 (193) adjusted/? 2 (tf 2 ) (194) perfect multicollinearity (197) dummy variable trap (201) imperfect multicollinearity (202) Exercises The first four exercises refer to the table of estimated regressions on page 206. 5.2 Using the regression results in column (1). and the R2 are measures of fit for the multiple regression model. 0 if male) Age = age (in years) Ntheast . .. When the four least squares assumptions in Key Concept 6. if Xx and X2 are highly correlated. computed using data for 1998 from the CPS. Explain why it is difficult to estimate precisely the partial effect of Xx.. The coefficients in multiple regression can be estimated by OLS. 0 if high school) Female ~ binary variable (1 if female. v 0 \ nv^ Review the Concepts 6.The highest educational achievement for each worker was either a high school diploma or a bachelor's degree. Will fa be an unbiased estimator of the effect on test scores of increasing the number of computers per student? Why or why not? If you think B-\ is biased.binary variable (1 if college. she regresses district average test scores on the number of computers per student.3 6... the OLS estimators are unbiased. Is age an important determinant of earnings? Explain. and number of children. and normally distributed in large samples.binary variable (1 if Region ~ South.1 Compute R2 for each of the regressions. 0 otherwise) South . For the purposes of these exercises. 0 otherwise) Midwest— binary variable (1 if Region = Midwest.4 are satisfied.

Suppose that a homeowner adds a new bathroom to her house.5.48 -2.S. Your critique should explain any problems with the proposed research and describe how the research plan might be improved.t (2) 5.62 0. the researcher collects salary and gender information for all of the firm's engineers.7' Critique each of the following proposed research plans. „„v„ „\ ytffctfk Vll^et fo-^-O ^ -VW A: (PQuyAll fft^P^ofS^ (in square feet). A researcher plans to study the causal effect of police on crime using data from a random sample of U.69 4." An estimated regression yields Prices 119.29 (3) 5.) Midwest (X5) y ^ A£ N f c * South (X6) Intercept Summary Statistics SER R z (1) 5 46 -2.090A#e 48. Suppose that a homeowner converts part of an existing family room in her house into a new bathroom. Lsize denote the lot size .62 0.yiDo there appear to be important regional differences? . (That is.002Lsize t.) Female [Xz) Age (A3) Northeast (X. What is the expected increase in the value of the house? b. Jennifer is a 28-year-old female college graduate from the Midwest. Use your answer to (a) and the expression for omitted variable bias given in Equation (6. BDR denote the number of bedrooms.c\Poor.44 2. A researcher is interested in determining whether a large aerospace firm is guilty of gender bias in setting wages. Include a discussion of any additional data that need to be collected and the appropriate statistical techniques for analyzing the data. To determine potential bias. Predict Sally's and Betsy's earnings. Betsy is a 34-year-old female college graduate.29 0. a. He plans to regress the county's crime rate on the (per capita) size of the county's police force.0.5 c.41. The researcher then plans to conduct a "difference in means" test to determine whether the average salary for women is significantly less than the average salary for men.6 c. Explain why this regression is likely to suffer from omitted variable bias. a.\56Hske + 0. Why is the regressor West omitted from the regression? What would happen if it was included? .75 6.176 6.1) to determine whether the regression will likely over. Compute the R2 for the regression.21 0.206 A -to * 5 A CHAPTER 6 Linear Regression with Multiple Regressors 1 Exercises 207 1. S a jjy 1S a 29-year-old female college graduate.27 / 6.27 0. which increases the size of the house by 100 square feet. Bath denote the number of bathrooms. and Poor denote a binary variable that is equal to 1 if the condition of the house is reported as "poor. SER . 12. do you think that fa > fa or fa < fal) 6.194 R2 n 4000 4000 4000 _C V 7i- • ^ %t 6.22 0. R2 = 0. Let Price denote the selling price (in \$1000). Age denote the age of the house (in years). What is the expected increase in the value of the house? Dependent variable: average hourly earnings (AHE). Data were collected from a random sample of 220 home sales from a community in 2003. Which variables would you add to the regression to control for important omitted variables? b. a.60 -0. counties. What is the loss in value if a homeowner lets his house run down so that its condition becomes "poor"? -\$.or underestimate the effect of police on the crime rate.2 + 0AS5BDR + 73ABath + 0.40 3. Hsize denote the size of the house (in square feet). sing the regression results in column (3): a. Juanita is a 28-year-old female college graduate from the South. Regressor College (#.190 6.72.69 0. Calculate the expected difference in earaifigs between Juanita and Jennifer.

A researcher is interested in determining whether time spent in prison has a permanent effect on a person's wage rate.91 (Yh X\i.-. b. . the causal effect of X\ on Y. var(u. in addition. Show that the least squares estimators satisfy fa = Y . b. [Hint: Look at Equation (6.X\)(X2i . it is best to leave out of the regression if it is correlated with Xxr 6. Run a regression of Course Eval on Beauty.X2) = 0. Assume that cor(Xi.11 (Requires calculus) Consider the regression model Y^faXu^faX2i¥Ul for / — 1 .n.i = l. the variance of 8 L is larger than it would be if X\ and X2 were Empirical Exercises 209 uncorrected..10 (Yj. The data set includes information on each person's current wage.5.X2i).t y ^ . Suppose that X\ and X2 are uncorrected. suppose that the model contains an intercept.Xu. and so on). Show that fa = S ^ A .. | Xu.208 CHAPTER 6 Linear Regression with Multiple Regressors b.faX2i + «. tenure (time in current job).1 Using the data set Teaching Ratings described in Empirical Exercises 4.faX2. This calculation was then repeated for people sleeping 6 hours. as well as whether the person was ever incarcerated. ethnicity. OneCredit. 6. Minority.. and AWEnglish. union status. What is the estimated effect of Beauty on Course_EvaH Does the regression in (a) suffer from important omitted variable bias? . would you recommend that Americans who sleep 9 hours per night consider reducing their sleep to 6 or 7 hours if they want to prolong their lives? Why or why not? Explain. .faX\ .17J in the Appendix 6. Also suppose that 2r=i(Ai. You are interested in fa. The researcher plans to estimate the effect of incarceration on wages by regressing wages on an indicator variable for incarceration. age. Specify the least squares function that is minimized by OLS. Suppose S/Li-Jq/Aa ^ 0. Compute the partial derivatives of the objective function with respect to hi and bic.0. He collects data on a random sample of people who have been out of prison for at least 15 years.2.4. occupation. As in (e). Suppose XUX-jXu = 0. (Notice that there is no constant term in the regression. Does this estimator suffer from omitted variable bias? Explain. e. X2i) satisfy the assumptions in Key Concept 6. a. Each survey respondent was tracked for 4 years. a. Show that jSi-2?-i(Ai. if you are interested in . 6. . . tenure.8 A recent study found that the death rate for people who sleep 6 to 7 hours per night is lower than the death rate for people who sleep 8 or more hours. Thus. education. Run a regression of Course_Eval on Beauty.4.2. carry out the following exercises. You estimate fa by regressing Y onto Xv (so that X2 is not included in the regression). A random sample of size n = 400 is drawn from the population. including in the regression the other potential determinants of wages (education. Comment on the following statements: "When X\ and X2 are correlated. 6.) Following analysis like that used in Appendix 4. f. X2i) = 4 and vai(Xu) .-XiXYi-YWZUiXuX. Female.i A * . In particular.2. How does this compare to the OLS estimator of B} from the regression that omits AV? Empirical Exercises E6. Assume that Xx and X2 are uncorrected. X2) .)2. c.6. Xu. and union status. n. d. Compute the variance of fa. including some additional variables to control for the type of course and professor characteristics.Derive an expression for fa as a function of thedata(Yi. He collects similar data on a random sample of people who have never served time in prison.. What is the estimated slope? b.. include as additional regressors Intro. . and so on. The death rate for people sleeping 7 hours was calculated as the ratio of the number of deaths over the span of the study among people sleeping 7 hours to the total number of survey respondents who slept 7 hours. The 1. Based on this summary. Compute the variance of Br. gender. a. X2i) satisfy the assumptions in Key Concept 6.1 million observations used for this study came from a random survey of Americans aged 30 to 102. Suppose that the model includes an intercept: Yj = fa + faXu 4.

carry out the following exercises.4.2 Using the data set CollegeDistance described in Empirical Exercise 4. b. R2 and R2. Rev_Coups. Explain why CueSO and Swmfg80 appear in the regression. RGDP60. Repeat (c) but now assume that the country's value for TradeShare is one standard deviation above the mean.5%.1) Derivation of Equation (6. In particular. Jim has the same characteristics as Bob except that his high school was 40 miles from the nearest college. d.3. Run a regression of ED on Dist.210 CHAPTER 6 Linear Regression with Multiple Regressors c. standard deviation. What is the estimated slope? b.1). Predict Jim's years of completed schooling using the regression in (b). His base-year composite test score (Bytest) was 58. What does this coefficient measure? f. Are the signs of their estimated coefficients (+• or . Predict Professor Smith's course evaluation. His family income in 1980 was \$26. a. but his father did not. Rev _Coups.16) yields Equation (6. He teaches a three-credit upper-division course. His high school was 20 miles from the nearest college. Ownhome. Is the estimated effect of Dist on ED in the regression in (b) substantively different from the regression in (a)? Based on this. Compare the fit of the regression in (a) and (b) using the regression standard errors. and Stwmfg80. carry out the following exercises. but include some additional regressors to control for characteristics of the student.A. E6. The unemployment rate in his county was 7. His mother attended college. Predict Bob's years of completed schooling using the regression in (b). does the regression in (a) seem to suffer from important omitted variable bias? d. Verify that the three-step process yields the same estimated coefficient for Beauty as that obtained in (b).3 Using the data set Growth described in Empirical Exercise 4. and his family owned a home. Construct a table that shows the sample mean.3 (the Frisch-Waugh theorem).1).3 states that I £i=/M (6.-. Oil. and the local labor market. (l/«)2.000. Run a regression of years of completed education (ED) on distance to the nearest college (Dist).75. The value of the coefficient on DadColl is positive.. Use the regression to predict the average annual growth rate for a country that has average values for all regressors. Why are the R2 and R2 so similar in regression (b)? t e.) = pXuo-uax. Assassinations. the student's family. Female.1) 211 h. E6. TradeShare.3. What is the estimated effect of Dist on ED'l c. Black. a. YearsSchool. but excluding the data for Malta. CueSO. d. Hispanic. include as additional regressors Bytest. Run a regression of Growth on TradeShare. Incomehi. Substitution of these limits into Equation (6.16) v\2 P crx and Under the last two assumptions in Key Concept 4.-=i(^. Estimate the coefficient on Beauty for the multiple regression model in (b) using the three-step process in Appendix 6. Assassinations and RGDP60. Bob is a black male.) what you would have believed 9 Interpret the magnitudes of these coefficients. Why is Oil omitted from the regression? What would happen if it were included? This appendix presents a derivation of the formula for omitted variable bias in Equation (6. Professor Smith is a black male with average beauty and is a native English speaker. 1 6. . Is it large or small in a real-world sense? c. What is the value of the coefficient on RevJZoupsI Interpret the value of this coefficient. DadColl. e. Equation (430) in Appendix 4.X (Vn)l!l=i(Xi-X)ut —^-» COV(M. YearsSchool. and minimum and maximum values for the series Growth.1 Derivation of Equation (6. g. and the state average manufacturing hourly wage was \$9. Include the appropriate units for all entries.

binary variable (1 if Region .) Northeast (JtT4) Midwest (X5) Soulh (Xb) Intercept Summary Statistics and Joint Tests ^-statistic for regional effects — 0 SER R • n 1 (1) 5.South. Is age an important determinant of earnings? Use an appropriate statistical test and/or confidence interval to explain your answer.05) 3. The highest educational achievement for each worker was either a high school diploma or a bachelor's degree.The worker's ages ranged from 25 to 34 years.27 0.48 (0. 7.190 4000 6. Regressor i)Uege(Ai) Feinale (X2) Aget-Y.69 (0-30) 0. The data set consists of information on 4000 full-time full-year workers.Northeast. Juanita is a 28-year-old female college graduate from the South.20) 0.69 (0.21 0.29 (0. Betsy is a 34-year-old female college graduate. f Dependent Variable: average hourly earnings (AHE).60 (0.75 (1.21) -2.2 Provide an example of a regression that arguably would have a high value of R2 but would produce biased and inconsistent estimators of the regression coefficients). 0 if high school) Female = binary variable (1 if female. 0 otherwise) 7.26) 12. 7.1 Add "*" (5%) and "**" (1%) to the table to indicate the statistical significance of the coefficients. and number of children. 0 otherwise) South .binary variable (1 if Region .2 Using the regression results in column (1): a. 0 otherwise) West .22 0.average hourly earnings (in 1998 dollars) College = binary variable (1 if college.04) 0.14) 4. Sally is a 29-year-old female college graduate. Why isn't the result of the joint test implied by the results of the first two tests? 7.176 4000 6.28) -0.3 Using the regression results in column (2): a. Jennifer is a 28-year-old female college graduate from the Midwest.10 6. b. Exercises The first six exercises refer to the table of estimated regressions on page 243. Explain why the OLS estimators would be biased and inconsistent.40 (1. .46 m\) 2.06) 6. Using the regression results in column (3): a. 0 otherwise) Midwest = binary variable (1 if Region = Midwest.27 (0. computed using data for 1998 from the CPS. i. let AHE . Construct a 95 % confidence interval for the difference in expected earnings between Juanita and Molly.62 (0. Construct a 95 % confidence interval for the expected difference between their earnings.29 (0.242 CHAPTER 7 Hypothesis Tests and Confidence Intervals in Multiple Regression hypothesis that fa = 0 and fa ~ 0. 0 if male) Age = age (in years) Ntheast — binary variable (1 if Region .20) 0.West. Is the male-female earnings difference estimated from this regression statistically significant at the 5% level? Construct a 95% confidence interval for the difference.20) (2) 5. Explain why the R2 is likely to be high.04) (3) 5.194 4000 b.21) -2. Is the college-high school earnings difference estimated from this regression statistically significant at the 5% level? Construct a 95% confidence interval of the difference. 7.44 (0.64 (0.The data set also contained information on the region of the country wlj^fe the person lived. marital status.4 Exercises 243 b. Molly is a 28-ycar-old female college graduate from the West.62 (0. Do there appear to be important regional differences? Use an appropriate hypothesis test to explain your answer. For the purposes of these exercises.

2. Consider the regression Yf = fa + B^Xy. Do you think that another scale might be more appropriate? Why or why not? e. a. SER = 5.8Poor.00048) f 0.311) (10.59Female +• 0 . and Xy denote a binary variable that equals 1 if the student is newly enrolled.6 Evaluate the following statement: "In all of tne regressions.08. Construct a 99% confident interval for the change in the value of her house.090Age . d.7 .13) and (7. the coefficient on Female is negative.S. 7.2 +. X2i) depends on X{> Is the OLS estimator of 8: unbiased 7. Construct the homoskedasticity-only ^-statistic for testing 03 ~ £ 4 = 0 in the regression shown in column (5).98) (0." Question 6.011) (0.3 to transform the regression so that you can use a /-statistic to test a.Y1( denote a binary variable that equals 1 if the student is assigned to a small class.77 + 5.5 (0. large.11 A school district undertakes an experiment to estimate the effect of class size on test scores in second grade classes. R2 ~. d.-. fa^fa. A homeowner purchases 2000 square feet from an adjacent lot. R2 = 0^ (0. (Hint: You must redefine the dependent variable in the regression. Consider the regression Yj . This provides strong statistical evidence of gender discrimination in the U. l56Hsize +• Q. I! Exercises 245 b. The district assigns 50% of its previous year's first graders to small second-grade classes (18 students per classroom) and 50% to regular-size classes (21 students per classroom). (Hint: What would happen if you included West and excluded Midwest from the regression?) 7. b.03) Comparing this regression to the regression for 1998 shown in column (2). converted into 1998 dollars using the consumer price index).18) (0.85. The results are AHE ^ 0. Is the coefficient on BDR statistically significantly different from zero? b. where a is a constant.72. Explain how you would construct a 95% confidence interval for the difference in expected earnings between Juanita and Jennifer.9) (2.1. The F-statistic for omitting BDR and Age from the regression is F— 0.61) (8. Let ^ d e n o t e the exam score for the /'th student.8 Referring to Table 7. Construct the R2 for each of the regressions. b. Typically five-bedroom houses sell for much more than two-bedroom houses. Is this consistent with your answer to (a) and with the regression more generally? c.23ABath + 0. SER = 41.-. Test fa — 84 = 0 in the regression shown in column (5) using the Bonferroni test discussed in Appendix 7. Is the statistic sig nificant at the 5% level? c.0. labor market. 4 0 4 ^ . each student is given a standardized exam. and statistically significant.244 CHAPTER 7 Hypothesis Tests and Confidence Intervals in Multiple Regression ii. Are the coefficients on BDR and Age statistically different from zero at the 10% level? ^ 7. Let Bi denote the causal effect on test scores of reducing class size from regular to small. Construct a 99% confidence interval for fa for the regression in column (5).5 reported the following regression (where standard errors have been added): Price •= 119. Do you think that E(Uj\X\j. At the end of the second-grade school year.1.5) a. + faX2l + «. was there a statistically significant change in thf coefficient on College? 7. Bx + a fa ~ 0.14) show two formulas for the homoskedasticity' only F-statistic. ' 7.ur Do you think that E(ut\X\j) ~ 0 ? Is the OLS estimator of B: unbiased and consistent? Explain. this time using data from 1992 (4000 observations selected at random from the March 1993 CPS.5 The regression shown in column (2) was estimated again. Students new to the district are handled differently: 20% are randomly assigned to small classes and 80% to regular-size classes.fa + 8{XU t.48. c.) 7. .485BDR J.10 Equations (7.2QCollege .20) (0. JSJ -t.002Lsize (23.1 in the text: a.94) (0. Show that the two formulas are equivalent. Lot size is measured in square feet. Use Approach #2 from Section 7.fa =.9 Consider the regression model Yj = fa ^-/B1 Xh + faX2i + «.U.

Run a regression of Growth on TradeShare. Is the coefficient statistically significant at the 5% level? b. a base specification (that includes a set of important control variables). Does the regression in (a) seem to suffer from omitted variable bias? d. Which do you think should be included in the regression? Using a table like Table 7. It has been argued that. Predict Bob's earnings using the estimated regression in (b). Alexis is a 30-year-old female worker with a college degree. Compare the fit of the regression in (a) and (b) using the regression standard errors. carry out the following exercises. Include a simple specification [constructed in (a)]. a. c. s / . Rev_Coups. carry out the following exercises. What is the estimated effect of Age on earnings? Construct a 95% confidence interval for the coefficient on Age in the regression. Other factors also affect how much college a person completes. Discuss how the estimated effect of Dist on ED changes across the specifications. Test the null hypothesis that both Female and Bachelor can be deleted from the regression. and several modifications of the base specification. Do you think that E(ui\Xu.1. Run a regression of average hourly earnings (AHE) on age (Age). Construct a 95% confidence interval for the effect of Beauty on Course_Eval.2 Using the data set TeachingRatings described in Empirical Exercise 4.3 to answer the following questions. An education advocacy group argues that. A regression will suffer from omitted variable bias when two conditions hold. examine the robustness of the confidence interval that you constructed in (a). Is this result consistent with the regressions that you constructed in part (b)? E7. Are the results from the regression in (b) substantively different from the results in (a) regarding the effects of Age and AHFf. What is the p-value of the /^statistic? Empirical Exercises E7. Assassinations. a.3 Use the data set CollegeDistance described in Empirical Exercise 4. c.4 Using the data set Growth described in Empirical Exercise 4. Bob is a 26-year-old male worker with a high school diploma. construct a table like Table 7. gender LEemale). Predict Alexis's earnings using the regression. What is a reasonable 95% confidence interval for the effect of Beauty on Course_EvaP. YearsSchool. and education (Bachelor). a.4.1. R2 and R2 Why are the R2 and R2 so similar in regression (b)? f.1 to answer the following questions. and RGDP60 can be omitted from the regression. What is the estimated intercept? What is the estimated slope? b. g. YearsSchool. b. Construct a 95% confidence interval for the coefficient on TradeShare. Consider the various control variables in the data set. Run a regression oiAHE on Age. on average. X2j) depends on X2? Will the OLS estimator of fa provide an unbiased and consistent estimate of the causal effect of transferring to a new school (that is. Run a regression of years of completed education (ED) on distance to the nearest college (Dist). e.2. Does controlling for these other factors change the estimated effect of distance on college years completed? To answer this question. blacks and Hispanics complete more college than whites. Empirical Exercises 247 a. taken as a group. Rev_Coups. E7. Assassinations.1 Use the data set CPS08 described in Empirical Exercise 4. Test the null hypothesis that Bachelor can be deleted from the regression. Run a regression of Course_Eval on Beauty.15 year if distance to the nearest college is decreased by 20 miles. Test whether. Is the advocacy groups' claim consistent with the estimated regression? Explain. controlling for other factors. b. What are these two conditions? Do these conditions seem to hold here? E7.246 CHAPTER 7 Hypothesis Tests and Confidence Intervals in Multiple Regression and consistent? Explain. a person's educational attainment would increase by approximately 0. but excluding the data for Malta. and RGDP60. Are gender and education determinants of earnings? Test the null hypothesis that Female can be deleted from the regression. being a newly-enrolled student)? Explain.

» Fh ' 8.4 0.2 Dependent variable: InWrice) Regressor Size \n(Size) ln(Size)2 Bedrooms Pool View Pool X View Condition Intercept Summary Statistics SER R 1 (1) 0.00042 (0.0078 (0.3 0. What will happen to the value of m if GDP increases by 2%? What will happen to m if the interest rate increases from 4% to 5%? You have estimated a linear regression model relating Y to X.14) 0.098 0.055) 8..69 (0.13 (0.071 (0.071 (0.~ 0. Suppose that in Exercise 8.60 (0.037 (0.12 (0. Bedrooms = number of bedrooms.037) 0. Repeat (a) assuming Sales2QW = 205. what is the expected change in price of building a 500-square-foot addition to a house? Construct a 95% confidence interval for the percentage change in pricey /-.296 CHAPTER 8 Nonlinear Regression Functions Review the Concepts ..73 .035) 6.036) 0. c. capital (K).032) 0.B^^GDP) f B2R.028) 0. Condition . ' ^ n> a.029) 0.036) 7.12 (0.fa -I.082 (0.Sales m M . where m is the quantity of (real) money.0036 (0.12 (0.57 (2. fat fa.60 (0.. How would you use regression analysis to estimate the production parameters? A standard "money demand" function used by macroeconomists has ffifrfprm ln(m) .03) 0.099 0. and fa are production parameters.026) 0. _ *.. but rather increased when K increased.2 you thought that the value of B2 was not constant. where A. Suppose that /3. Explain how you would specify a nonlinear regression to model this shape.40) 8.73 0.72 0.binary variable (1 if house has a nice view. Variable definitions: Price = sale price (\$). and an error term u using the equation Q . 0 otherwise): View .0022 (0. Suppose that you have data on production and the factors of production from a random sample of firms with thV-sanie Cobb--Douglas production function. 0 otherwise). Can you think of an economic relationship with a shape like this? A "Cobb-Douglas" production function relates production (Q) to factors of production.50) 0.087) 0.. ' 0.102 0. Using the results in column (1). . .5 Exercises 7 ieo • --' iCC fh> .XKplL^2M^eu. Compute the percentage increase in sales using the usual formula (5flte.045) 10. "I think that the relationship between Y and X is nonlinear.0 and fa . Size = house size (in square feet).071 (0.02.53) 0. O ^ T O K O ^ O A / tit. 8.39) 0. How could you use an interaction term to capture this effect? Exercises 297 8. oQpAr2 (j&S) ^ ^ C \ _ ^ v i & y ^ t f W fo^to^S juEfy'*S E r _^s°l-^?5)-5-1 8. •f a.035) 0.030) 0.69 (0.02 (7.029) 0..binary variable (1 if real estate agent reports house is in excellent condition.035) 6. How good is the approximation when the change is small? Does the quality of the approximation deteriorate as the percentage change increases? QJ± ^ ^ ™ o ^ ^ e ^ C .68 (0.12 (0. = 1.026 (0.027 (0. GDP is the value of (real) gross domestic product. 0 otherwise).035) 6.034) 0. ) mii . . Sales20m — 250.054) 0. and Sales2Q\(\ = 500." Explain how you would test the adequacy of your linear regression.000038) (2) (3) (4) (5) 8.099 0.1 Sketch a regression function that is increasing (has a positive slope) and is steep for small values of X but less steep for large values of X. labor (/7h and raw materials (M)..10) 0. Compare this value to the approximation 100 X "« S f e s ? o t 9 i 100 X [ln(Sales2Qio) .97 (0 069) 0. J . and R is the value of the nominal interest rate measured in percent per year.2/ Suppose that a researcher collects data on houses that have sold in a particular neighborhood over the past year and obtains the regression results in the table shown below.027 (0.1N Sales in a company are \$196 million in 2009 and increase to \$198 million in 2010.74 0.099 0. Your professor says.ln(Sa/es2oo9)]b.027 (0.63 (0. Pool = binary variable (1 if house has a swimming pool. ^ .73 0.034) 0.071 (0..

Describe a nonlinear specification that can be used to model this form of nonlinearity.28) interprets the coefficient on interacted binary variables using the conditional mean zero assumption.3? b. Describe a nonlinear specification that can be used to model this form of nonlinearity. F = 2. (Hint: For the log-log model assume that u and X are independent as is done in Appendix 8. A y/AX> = fa + faX: (effect of change in X2 holding X± constant).03) (0. Explain what this value means.8). The estimated coefficient on Female is -0.05) i. Sketch the following regression functions (with values of X between 5 and 100 on the horizontal axis and values of Y on the vertical axis): a.2 for the linear and log-log models. ii.345.0xln(^T) f 4 . hi. are added to the regression: bi(Earnings) .8 Exercises 301 i. public corporations in the 1990s. Let Female be an indicator variable that is equal to 1 for females and 0 for males. i.12 The discussion following Equation (8. SER = 2. 0 1 ^ .48 -. ii.] + faXu + faX2i -f fa(Xu X X2i) +».004) (0. c. -f (fa f faX. (0.0Z.65.0xln(A') f-4. in millions of dollars) and stock return (a measure of firm performance. but with Z — 0.86 0. Explain what this value means.2 . r = 2. 0 .01) (0. 0 + 3.faX2)AX. w i t h Z . in percentage points). How would you test whether the researcher's conjecture was better than the linear specification in column (7) of Table 8. f = 1 . with Z = 1.04) (0. This . 0 Z .0xin(AT). 8. b. R2 = 0. r . The SER is 2.2 for the log-linear model. A researcher suspects that the effect oT income on test scores is different in districts with small classes than in districts with large classes. A regression of the logarithm of earnings onto Female yields ]r7(Ea7nings) . The study compares total compensation among top executives in a large set of U. The coefficient on In(MarketValue) is 0. Does this regression suggest that female top executives earn less than top male executives? Explain. i.7 This problem is inspired by a study of the "gender gap" in earnings in top corporate jobs [Bertrand and Hallock (2001)]. 0 ^ . Same as (i). ii.3. e.Y).) 8. Explain why it has changed from the regression in (a). (Each year these publicly traded corporations must report total compensation levels for their top five executives.371n(MarketValue) + O. 8. Explain what this value means. c.300 CHAPTER 8 Nonlinear Regression Functions i. "I"wo new variables.0A4Female. ^Y|^XX = fa + B3X2 (effect of change in XY holding X2 constant).2 . (0.9 Explain how you would use "Approach #2" of Section 7.)AX2 + faAX^X2.46. See Exercise (7. c.65.003) n .. 0 X Z X In(.1 . Same as (1).1. but with Z = 0. 1 b. d. How would you test whether the researcher's conjecture was better than the linear specification in column (7) of Table 8T3? 8.0 .37.28.0 + 3 .28Female + 0. the market value of the firm (a measure of firm size. The coefficient on Female is now -0.) a.6. Z is a binary variable. If Xt changes by AA^ and X2 changes by AX2.S.10 Consider the regression model Yt^fa Use Key Concept 8. 8.^ 1 2 5 .9). X i s a continuous variable that takes on values between 5 and 100. then A Y ~ (fa . 8.3 to calculate the confidence interval discussed below Equation (8.QOAReturn. 0 x l n ( ^ ) .0-3. 0 + 3.44. f .670.1 to show: a. Does this regression suggest that there is gender discrimination? Explain. i.11 Derive the expressions for the elasticities given in Appendix 8. iv. Are large firms more likely to have female top executives than small firms? Explain. ii. ii. b. [Hint:This requires estimating a new regression using a different definition of the regressors and the dependent variable. ii.

on \n(Age).E(Yj\X^= 1. carry out the following exercises. that 8t ~. Use the conditional mean independence assumption E(UJ\X\J. gender (Female). What does the regression predict for her value of \n(AHE)1 What is the predicted difference between Alexis's and Jane's earnings? Bob is a 30-year-old male with a bachelor's degree. how are earnings expected to change? If Age increases from 33 to 34. X2i = 1) . Is there evidence that Age has a nonlinear effect on Course_Eval? Is there evidence that Age has any effect on Course EvaH Empirical Exercises E8. After running all these regressions (and any others that you want to run). Suppose that you estimate the interaction regression in (c) using the combined data and that E(ut\X^X2iy-**E(uj\X2i). summarize the effect of age on earnings for young workers. Is the effect of Age on earnings different for men than for women? Specify and estimate a regression that you can use to answer this question. d. Show that 81 is the class size effect for new students. Female.. ln(. b. Do you prefer the regression in (d) to the regression in (b)? Explain. k. that is. a. Age2. X2j) = E(Uj\X2j) to show that fa = yh By + B3 = 8U and fa . Suppose that you estimate the regression Yj — 80 f S^Xn + «.using only the data on returning students.\XU = 0. how are earnings expected to change? c.X2l-§)Explain why yy is an unbiased estimator of y^. Bachelor. Female. Do you prefer the regression in (d) to the regression in (c)? Explain. Female. Age2. = E(Yi\Xli^ 1. and the interaction term Female X Bachelor. Female. UAge increases from 25 to 26. Run a regression of the logarithm of average hourly earnings. how are earnings expected to change? If Age increases from 33 to 34. Female. Is the effect of Age on earnings different for high school graduates than for college graduates? Specify and estimate a regression that you can use to answer this question. \n(AHE). Show that Bj and fa are unbiased but that fa '1S m general biased. (c).using only the data on new students. Consider the hypothetical experiment in Exercise 7. how are earnings expected to change? d. Empirical Exercises 303 25 to 26.11. Consider the regression for both retijrning and new students. f. If Age increases from 25 to 26. and Bachelor. how are earnings expected to change? e.1 Use the data set CPS08 described in Empirical Exercise 4. X^-O) E{Yi\Xu-'-{). Suppose that you estimate the regression Yt — yQ +. and education (Bachelor). Yj •= fa + faXu +.yx (the difference in the class size effects).E(Y. on Age. Explain why 6j is an unbiased estimator of 8y c. on Age. Show that y 1 is the class size effect for returning students. Run a regression of the logarithm of average hourly earnings. X2i = 1). and Bachelor. Intro.. What does the coefficient on the interaction term measure? Alexis is a 30-year-old female with a bachelor's degree. on Age. OneCredit. and (d) for males with a high school diploma. Would your answer change if you plotted the regression function for females with college degrees? i.faX2j + fa(Xu X Xtf + Uj.2 Using the data set TeachingRatings described in Empirical Exercise 4. and NNEnglish. \n(AHE). a. h. Plot the regression relation between Age and ln(AHE) from (b). a.y\XXi + «. What does the regression predict for his value of \n(AHE)7 Jim is a 30-year-old male with a high school degree.302 CHAPTER 8 Nonlinear Regres^ion-FwflGtions exercise shows that interpretation also applies under conditional mean independence. What does the regression predict for his value of in(AHE)'? What is the predicted difference between Bob's and Jim's earnings? j. Describe the similarities and differences between the estimated regression functions. Run a regression of the logarithm of average hourly earnings. how are earnings expected to change? If Age increases from 33 to 34. E8. Add Age and Age2 to the regression. If Age increases from 25 to 26. that is. and Bachelor. Estimate a regression of Course_Eval on Beauty. how are earnings expected to change? b. that y. Run a regression of \n(AHE). Do you prefer the regression in (c) to the regression in (b)? Explain.4//£). UAge increases from & i&tt . how are earnings expected to change? If Age increases from 33 to 34.1 to answer the following questions.Si . Run a regression of average hourly earnings (AHE) on age (Age). What does the regression predict for her value of \n(AHE)'l Jane is a 30-year-old female with a high school degree. g. b.2. 1. Minority.

Although such regression functions cannot be estimated by OLS.) 4 "'" The negative exponential growth function is graphed in Figure 8.5 to 1.1 Regression Functions That A r e N o n l i n e a r in t h e Parameters (a) A logistic curve (b) A negative exponential growth curve The nonlinear regression functions considered in Sections 8.38) Negative exponential growth. how ever. Because they are linear in the unknown parameters. and you have data on n industries.2 and 8. The logistic function smoothly increases from a minimum of 0 to a maximum of l. an asymptote as income increases to infinity). the value of the function is nearly 0 and the slope is flat. Use regression (5) to predict the increase in Growth. Using regression (5). The logarithmic specification has a positive slope for all values of income. the function approaches 1 and the slope is flat again. a single independent variable A describes an industry characteristic. as income gels very large.2 to model the relation between test scores and income have some deficiencies.38). economic reasoning leads to regression functions that are not linear in the parameters.39) Y > ~ i + e. As can be seen in the graph. and an asymptote at /30 as X tends to infinity.Thc logistic regression model with a single X is (8. The negative exponential growth regression model is Yj = 80[l-e^x> ft)]+U(. which is implausible. Logistic curve. This family of nonlinear regression functions is both rich and convenient to use. but as X increases it reaches an asymptote of fa. however.306 CHAPTER 8 Nonlinear Regression Functions e. such as the adoption of database management software in different industries. The functions used in Section 8. The dependent variable is between 0 (no adopters) and 1 (100% adoption). those parameters can be estimated by OLS after defining new regressors that are nonlinear transformations of the original Xs. Functions That A r e Nonlinear in the Parameters We begin with two examples of functions that are nonlinear in the parameters.3 are nonlinear functions of the A"s but are linear functioris of the unknown parameters. The logistic function with a single ^Tis graphed in Figure 8. The negative exponential growth model provides a nonlinear specification that has a positive slope for all values of income. Use regression (3) to predict the increase in Growth. the polynomial models can produce a negative slope for some values of income.<£+&*. it makes sense to use instead a function that produces predicted values between 0 and 1. The dependent variable is the fraction of firms in the industry that have adopted the software. so for some incomes the predicted value for a district will exceed the maximum possible score on the test. a country contemplated a trade policy that would increase the average value of TradeShare from 0. and has an upper bound (that is. which has predicted values that lie between 0 and 1. (8.39). . Because a linear regression model could produce predicted values less than 0 or greater than 1. the curve is steeper for moderate values of X. they can be estimated using an extension of OLS called nonlinear least squares. For small values of X. The slope is steep for low values of X. Part (b) plots the negative exponential growth function of Equation (8. the logistic function has an elongated "S" shape. For example. is there evidence of a nonlinear relationship between TradeShare and Growth"? f. the predicted values increase without bound. We then provide a general formulation.12b. Suppose that you are studying the market penetration of a technology. has a slope that is greatest at low values of income and decreases as income rises. which has a slope that is always positive and decreases as X increases. Part (a) plots the logistic function of Equation C8. In some applications.12a. In 1960. and for large values of X. \mi\xni\Xmmm Regression Functions That Are Nonlinear in the Parameters Two Functions That Are Nohlinear in Their Parameters 307 %A¥pmmx 8.

one or more of the regressors are measured with error. or when the error term is correlated across different observations.3 Economic variables are often measured with error. Which should the researcher use? Why? 9. When regression models are used solely for forecasting. OLS estimators will be biased and inconsistent if the regressors and error terms are correlated. A researcher estimates the effect on crime rates of spending on police by using city-level data. however. Standard errors are incorrect when the errors are heteroskedastic and the computer software uses the homoskedasticity-only standard errors. b.for all values of i and. Suppose that Yj is measured with error. Consider the population regression Y] — fa + B\Xj -f vh where v. "Measurement error in the Xs is a serious problem. Does this mean that regression analysis is unreliable? Explain. confidence intervals and hypothesis tests are not valid when the standard errors are incorrect. 9.i.satisfies the assumptions in Key Concept 4. Measurement error in Kis not. f wt. The first uses the homoskedasticity-only formula for standard errors. the sample is chosen nonrandomly from the population. Show that v. Explain how sample selection bias might invalidate the results. Explain how simultaneous causality might invalidate the results. so the data are Yl=Yi + w„ where wt is the measurement error which is i. is the regression error using the mismeasured dependent variable. .2 describes the problem of variable selection in terms of a trade-off between bias and variance. there are two types of threats to internal validity.1 9. 3.w. Show that the regression Yt — fa -*. Can confidence intervals be constructed in the usual way? \ 93 e.is independent of Yj and X}.2 What is the difference between internal and external validity? Between the population studied and the population of interest? Key Concept 9. Use the concept of external validity to determine if these results are likely to apply to Boston in the 1970s.3. that the regression model be externally valid for the forecasting application at hand." Labor economists studying the determinants of women's earnings discovered a puzzling empirical result. an incorrect functional form is used. they Key Terms population studied (312) internal validity (313) external validity (313) population of interest (313) functional form misspecification (318) Review the Concepts 9. What is this trade-off? Why could including an additional regressor decrease bias? Increase variance? . A researcher estimates a regression using two different software packages. The second uses the heteroskedasticity-robust formula. Second.. First. Are the OLS estimators consistent? d. A study is internally valid if the statistical inferences about causal effects are valid for the population being studied. 5. s y Consider the one-variable regression model Yj — fa + faX-t -I.) c.5 9. It is critical. A study is externally valid if its inferences and conclusions can be generalized from the population and setting studied to other populations and settings. a. (Assume that iv(. 2. or there is simultaneous causality between the regressors and dependent variables.d. and has a finite fourth moment. Statistical studies are evaluated by asking whether the analysis is internally and externally valid. Los Angeles in the 1970s. m errors-in-variables bias (319) classical measurement error model (320) sample selection bias (323) simultaneous causality (324) simultaneous equations bias (325) Exercises 341 9. the study concluded that advertising on buses and subways was more effective than print advertising.\$xXt + v.340 CHAPTER 9 Assessing Studies Based on Multiple Regression Summary 1.3. 4.ut and suppose that it satisfies the assumption in Key Concept 4. Evaluate these statements.6 Exercises y^v^0"^-* > \j>k t& ^ 9. Regressors and error terms may be correlated when there are omitted variables. New York in 2010. and independent of Yj and Xj.V Suppose that you have just read a careful statistical study of the effect of advertising on the demand for cigarettes. Using data from New York during the 1970s. Y(.4 Suppose that a state offered voluntary standardized tests to all its third graders and that these data were used in a study of class size on student performance. The standard errors are very different. Using randomly selected employed women. In regression estimation of causal effects. it is not necessary for the regression coefficients to be unbiased estimates of causal effects.

but data on a related variable Zj are available.11 Read the box "The Demand for Economics Journals" in Section 8. Discuss the internal and external validity of the estimated effect of price per citation on subscriptions.l^^^vwS*^^ Exercises 343 Another researcher is interested in the same regression. a." ^~^^ b.] 9. and <2.2. They found that women with more children had higher wages.2) 9. and so forth).1) (12.7\ Are the following statements true or false? Explain your answer.10 Read the box "The Return to Education and the Gender Gap" in Section 8. Use these to determine the regression statistics.) Suppose that the sample is very large.100 i. education. and u denotes factors other than price that determine demand. Is the estimated slope too large or too small? (Hint: Remember that demand curves slope down and supply curves slope up.1.12 Consider the one-variable regression model Yt — fa + B^Xj + ut and suppose that it satisfies the least squares assumptions in Key Concept 4.y0 + yi_P + v. A random sample of observations of (Qj. c.5) be useful for predicting test scores in a school district in Massachusetts? Why or why not? Consider the linear regression of TestScore on Income shown in Figure 8. (15.) 9. observation 2 entered twice.8X. Pj) is collected. Qj is the regressand and Pj is the regressor. and are mutually uncorrelated.8 9. 9. . Derive the variance of P.3." •V'-^' Would the regression in Equation (9. so he has 200 observations (with observation 1 entered twice.1 + 66. [Hint: Use Equations (4.] ii. where Q denotes quantity.18).81. a. and the value of Xt is estimated using Xj — E(X^Zj). b.) [This empirical puzzle motivated James Heckman's research on sample selection that led to his 2000 Nobel Prize in economics. The regressor Xt is missing. d. Supply for the commodity is given by Q . P denotes price. 9.9 oft. observations for (YhXj) yield the following regression results: Y . The demand for a commodity is given by Q .K 2 = ^ _ . Use your answers to (b) and (c) to derive values of the regression coefficients. a. "Each of the five primary threats to internal validity implies that X is correlated with the error term.) Y = ( X.Xt. and so forth).32. variances and covariances of Y and X as functions of the "correct" values. Would either of these regressions provide a reliable estimate of the effect of income on test scores? Would either of these regressions provide a reliable method for forecasting test scores? Explain. A researcher uses the slope of this regression as an estimate of the slope of the demand function (/3i). SER = 15. (Hint: Notice tnat w o m e n w n o l Vrvrf ^° n o t WOT k outside the home are missing from the sample.3 and column (2) of Table 9. Explain how sample selection might be the cause of this result.3. "An ordinary least squares regression of Y onto X will be internally inconsistent if X is correlated with the error term. 9. controlling for these other factors.342 CHAPTER 9 Assessing Studies Based on Multiple Regression regressed earnings on the women's number of children and a set of control variables (age. SER = )( —) -. I. Suppose that u and v both have a mean of zero. Derive the means of P and £).fa + BAP •+• u.6 Suppose that n .d.i. where v denotes factors other than price that determine supply.8).4 Using the regressions shown in column (2) of Table 8. is regressed on Pr (That is. R2 = 0.5 b. and the covariance between Q and P. See Heckman (1974). Let wt = X.3 to compare the estimated effects of a 10% increase in district income on test scores in California and Massachusetts. . have variances a2 and a2. but he makes an error when he enters the data into his regression program: He enters each observation twice. the variance of Q. construct a table like Table 9. Discuss the internal and external validity of the estimated effect of education on earnings. Which (if any) of the internal validity conditions are violated? 9.3. occupation. Using these 200 observations. Solve the two simultaneous equations to show how Q and P depend on u and v. what results will be produced by his regression program? (Hint: Write the "incorrect" values of the sample means.2 and the nonlinear regression in Equation (8. 9.7) and (4.

Regression with entity fixed effects controls for unobserved variables that differ from one entity to the next but remain constant over time. 7. The researcher is interested in the effect of education on earnings. 6.1. Time fixed effects control for unobserved variables that are the same across entities but vary over lime. b. Use the results in column (4) to predict the change in the number of traffic fatalities in the next year.370 CHAPTER to Regression with Panel Data Summary 1. Use the results in column (4) to predict the change in the number of traffic fatalities in the next year. A regression with time and entity fixed effects can be estimated by including binary variables for n . The drinking age in New Jersey is 21. to describe panel data? What does / refer to? What does / refer to? entity and time fixed effects regression model (360) autocorrelated (363) serially correlated (363) heteroskedasticity. ^ w * ^ . In panel data. Standard errors need to allow both for this autocorrelation and for potential heteroskedasticity. New Jersey has a population of 8.2 Exercises 371 A researcher is using a panel data set o n « = 1000 workers over T= 10 years (from 2001 through 2010) that contains the workers' earnings.4 Exercises 10.~w)r "j U» ^X^-V"*-* x tion of the regression and the statistical test you would use. explain why the regression error for a given individual might be serially correlated.2. Suppose that New Jersey lowered its drinking age to 18. In the context of the regression you suggested for Question 10. Can you think of examples of time-specific variables that might be correlated with education and earnings? How would you control for these person-specific and time-specific effects in a panel data regression? Can the regression that you suggested in response to Question 10. Entity fixed effects regression can be estimated by including binary variables for n . correlated over time within an entity.St #i. c Suppose that real income per capita in New Jersey increases by 1% in the next year. that is. 4. gender.and autocorr elation-co nsistent (HAC) standard errors (364) clustered standard errors (364) h * -t <i i . Panel data consist of observations on multiple (n) entities—states.3 10. education. Should time effects be included in the regression? Why or why not 9 v ^ e.1 entities and binary variables fox T . fixed effect regression can be estimated by a "before and after" regression of the change in Y from the first period to the second on the corresponding change in X.1 This exercise refers to the drunk driving panel data regression summarized in Table 10. a. When there are two time periods.. Construct a 95% confidence interval for your answer. Give some examples of unobserved person-specific variables that are correlated with both education and earnings.1 million people. Use the results in column (4) to predict the number of lives that would be saved 3&t. •O^*?6TM.* ft ^tUrX* M\ fidence interval for your answer. How would you test this hypothesis? (Be specific about the specifica+ Key Terms panel data (348) balanced panel (348) unbalanced panel (348) fixed effects regression model (354) entity fixed effects (354) time fixed effects regression model (359) time fixed effects (359) Review the Concepts 10. Suppose that New "£Ao* T Vvrtt^ON/^ j e r s e y i n c r eased the tax on a case of beer by \$1 (in 1988 dollars). and age. 5.1 Why is it necessary to use two subscripts. variables are typically autocorrelated. 3. 10. and one way to do so is to use clustered standard errors.1 entities plus the observable independent variables (the X's) and an intercept. Construct a 90% conv. 10. i and f.. 2.) . people.1 time periods plus the Xs and an intercept.\^»» WO over the next year. firms.2 be used to estimate the effect of gender on an individual's earnings? Can that regression be used to estimate the effect of the national unemployment rate on an individual's earnings? Explain. and so forth —where each entity is observed at two or more time periods (T). Construct a 95% confidence interval for your answer. A researcher conjectures that the unemployment rale has a different ^ \ v effect on traffic fatalities in the western states than in the other states.

Will this regression give reliable estimates of the effects of the regressors (age. and CXJ + Xjt is an unobserved individualspecific time trend.1 Some U.372 CHAPTER 10 Regression with Panel Data 10.0C^ >> Q v where 1 ~ 1.5. are the fixed entity effects. What will happen if you try to estimate the coefficients of the regression by OLS? \ 10. do you think that the estimated values of a.S.vis) = 0 for / ^ s in Equation (10.11). let Yit = fa + faXit + y. Show the result in (a) for general n. A r )? A Do the fixed effects regression assumptions in Key Concept 10." b. — ctn. + a. The researcher collects data on the snowfall in each state for each year in the sample (Snowit) and adds this regressor to the regressions. a„ consistently estimated as n * 00 with T fixed? (Hint: Analyze the model with no X's: Ylt — a-t + uit. 8T> y2. 82.22). express one of the variables Dl ( .faXUt + 82B2C + • • • + 8TBTt + y2D2j + • • • . A researcher believes that traffic fatalities increase when roads are icy and so states with more snow will have more fatalities than other states. • . Entity t i n time period 3? p „ •' c Entity 3 in time period 1? r x * ^ R '*&? ' ^ P. if T = 2..7 Empirical Exercises E10. a. states have enacted laws that allow citizens to carry concealed weapons. Suppose that n — 3.22) to show that XiX = 4 ( ^ 2 " Xa) and Xa ~.5 Consider the model with a single regressor Yj. d.XjX)2]. so thatjBf* = \"ZU(Xa .2 Consider the binary variable version of the fixed effects model in Equation (10. and so forth. Using the regression in Equation (10.)..28)? Explain. that is. and previous year's earnings) on earnings? Explain.8 Empirical Exercises 373 ^ ^ ? > > w a.9 / 10. Apply this list to the empirical analysis in Section 10. These laws are known as "shall-issue" laws because they instruct local authorities to issue a concealed weapons permit to all applicants who are citizens. How are the coefficients (fa.at + ult. and the worker's earnings in the previous year using fixed effects regression.) 10. Entity 1 in time period 1? &«* . Xtt) from the linear panel data model Yjr^Xufa+aj+kjt + Ujt. union status... Dl t .. D\ and XQJ[ as a perfect linear function of the others. + uit.. are mentally competent.) 10. (Hint: Check the fixed effects regression assumptions in Section 10. D2{ — 1 if / = 2 and 0 otherwise.10 In a study of the effect on earnings of education using panel data on annual earnings for a large number of workers. b. How would you estimate 8{l a.3 / 10. A 1 ? . education.4 Section 9.« — 1 for all i.This model also can be written as Yjj = fa +. T^ 4). where Xn. except with an additional regressor. union status.y2D2t + •••+..ynDn. Show that. before Equation (10. = 1 if t = 2 and 0 otherwise. f «((. The researcher collects data on the average snowfall for each state and adds this regressor (AverageSnowJ to the regressions given in Table 10. what is the slope and intercept for a. Consider observations (Yit.6 and thereby draw conclusions about its internal validity.\(Xa . If n is large (say.Yil)]/[XU(Xj2 . t. and have not been convicted of a felony .Xj. D2j.are approximately normally distributed? Why or why not? (Hint: Analyze the model Ylt . n = 2000) but Tis small (say.) b.11 Let faM denote the entity-demeaned estimator given in Equation (10. where B2. c. yn) related to the * coefficients (a\. — B^Xij..Xn)(Yj2 . and let faA denote the "before and after" estimator without an intercept.N. T. ir = 1 . that is. In the fixed effects regression model..H).Dlt +.y„Dnt + uit.2 gave a list of five potential threats to the internal validity of a regression study.f A.6 10. 0?M = (3?A. education.3 imply that cov( vlt. Entity 3 in time period 3? BB *"U^ / £ « 10. b. Show that the binary regressors and the "constant" regressor are perfectly multicollinear.] ^ ' 10. [Hint: Use the definition of x). Comment on the following methods designed to estimate the effect of snow on fatalities: 10.1. a researcher regresses earnings in a given year on age.

1999. b. (ii) workers affected by workplace smoking bans. a. c.7).pearsonhighered. Estimate a probit model using the same regressors as in Empirical Exercise 11.p — q. to answer the following: ^"^ C^ oi' # * * a. and (hi) workers not affected by workplace smoking bans. where 411 faXt)[l (fa + P&)]. a. A study explaining the number of minutes that a person spends talking on a cell phone during the month? b.Y2 Yn. hsgrad. colsome. or generic cola? d.747. A study of the number of cell phones owned by a family? Empirical Exercises £11. age.1(c). Test the hypothesis that the probability of smoking does not depend on the level of education in the regression in (c). a. b. and other individual characteristics. black.000 U.410 CHAPTER 11 Regression with a Binary Dependent Variable 11. a. apply for a mortgage. Based on the regression in (c). and how would it bias the results? 11. In this assignment you will estimate the effect of workplace smoking bans on smoking using data on a ii. d. Construct a 95% confidence interval for your answer to (a). Two applicants. Is ut heteroskedastic? Explain. . one white and one black.11 (Requires Appendix 11. colgrad.89(4): 728.(j30 + Equation (2.They have the same values for all the regressors other than race.1 It has been conjectured that workplace smoking bans induce smokers to quit by reducing their opportunities to smoke. e. A study explaining grades (A through F) in a large Principles of Economics class? c. age2. ) .9 \ Use the estimated linear probability model shown in column (1) of Table 11. c Estimate a linear probability model with smoker as the dependent variable and the following regressors: smkban. "Do Workplace Smoking Bans Reduce Smoking?" American Economic Review. non-Hispanic male college graduate with no workplace smoking ban. based on the substance of this regression. indoor workers from 1991 to 1993. hsdrop. available on the textbook Web site www. and hispanic.com/stock_watson in the file Smoking. 11.3) Derive the likelihood function. (Requires Section 11. b.10 (Requires Section 11. [Hint: Review sample of 10. explaining the change in the estimated effect of a smoking ban between (b) and (c). Show that E(UJ\XJ) = 0. Test the hypothesis that the coefficient on smkban is zero in the population version of the regression in (c) against the alternative that it is nonzero. female.p . and the random variable's are denoted Yl. is there a nonlinear relationship between age and the probability of smoking? Plot the relationship between the probability of smoking and age for 18 < age ^ 65 for a white. Pepsi.] c.1. Estimate the probability of smoking for (i) all workers. Suggest a reason. Ell. Derive the likelihood function for the parameters p and q.2 This exercise uses the same data as Empirical Exercise 11. at the 5% significance level. Think of an important omitted variable that might bias the answer in (a). The data set contains information on whether individuals were or were not subject to a workplace smoking ban. and Pr( Y = 3) _ 1 .3 and calculus) Suppose that a random variable Khas the following probability distribution: Pr( Y = 1) . How much more likely is the black applicant to be denied a mortgage? b. Fr(Y = 2) = q.Description. Show that v a r f u ^ .S. Derive formulas for the MLE of p and q. b. What is it. Test the hypothesis that the coefficient on smkban is zero in the population version of this probit regression against the alternative that it is 'These data were provided by Professor William Evans of ihe University of Maryland and were used in bis paper with Matthew Farrelly and Edward Montgomery.fa faXf4 Uj. Compare the estimated effect of a smoking ban from this regression with your answer from (b). A random sample of size n is drawn from this distribution. d. What is the difference in the probability of smoking between workers affected by a workplace smoking ban and workers not affected by a workplace smoking ban? Use a linear probability model to determine whether this difference is statistically significant. available on the Web site. r ^ \^H.3) Which model would you use for: a. A study of consumers' choices for Coke. whether the individuals smoked.2 1 Empirical Exercises Yt . 7 A detailed description is given in Smoking_.8 Consider the linear probability model Yr(Yj = l\Xj)-fa^faXj. Does the probability of smoking increase or decrease with the level of education? f.

3 is satisfied with Z( = Xr b.Exercises 456 CHAPTER 1 2 Instrumental Variables Regression 12.•. Then the predicted value from the first-stage regression is Xj = 7j-0 + TTIZJ. 12. c Suppose that the recession lasts less than 1 year. I s ^ . How would you estimate fa.0 ^ L S 12. a.l) This question refers to the panel data regressions summarized in Table 12. a. where u] and uf are mutually independent i. c.14).6 (In an instrumental variable regression model with one regressor. Suppose that the first three assumptions in Key Concept 12.and Z2i. Does this suggest that E(u-\ZU) ^ 0? Explain. + it*. If the current price per pack is \$7. and Zt is an instrument. Zit the regression of Xj onto Zt has R2 ~ 0. Z2i) ^ 0? Explain. show that Key Concept 12.4 are satisfied with this choice of Z-P c.4 In their study of the effectiveness of cardiac catheterization. Show that Xt is a valid instrument. b. use the regression in column (1) to predict the change in demand. Show that the TV estimator constructed using Z. Would the regression provide a reliable answer to the question posed in (a)? Why or why not? 12. where X.500? 12. is independent of ( Y h Xt. Suppose that she uses the estimator from the second-stage regression of TSLS: £2 = „ 12 2 " i(*r . Use the definition of the sample variance and covariance to show that s^Y TTISZY and \$ | = TT\S2Z. ( ^ .50 per pack. and a market equilibrium condition Qst — Q?. Z]f.= Xt is identical to the OLS estimator. a. Consider TSLS estimation with a single included endogenous variable and a single instrument.6.2 Consider the regression model with a single regressor: Yj . random variables. B\. Zj = Xi? V12. W. Consider a product market with a supply function Q] = fa + /3.4 x . Suppose that the United States enters a recession and income falls by 2%. the value of the . Show that Pt and uf are correlated. Show that the IV regression assumptions in Key Concept 12.50. Xr and \ J two instruments.That is.1). i .. Do you think that the regression in column (1) will provide a reliable answer to the question in (b)? Why or why not? d» Suppose that the F-statistic m column (1) was 3. Use this result to fill in the steps of the derivation in Appendix 12. a demand function Qf — y0 + "f. is correlated with u. McClelian.) b.4). c.where Xt is the fitted value from the first-stage regression. Show that the OLS estimator of B t is inconsistent.P. and -y0? . Is this estimator consistent? (For 12. Use the regression in column (I) to predict the change in demand. Suppose that the federal government is considering a new tax on cigarettes that is estimated to increase the retail price by \$0.6 instead of 33.0 o S " MSLSK'f. both with a mean of zero. Xr and one instrument.3 A classmate is interested in estimating the variance of the error term in Equation (12. Is Zj a strong instrument? [Hint: See Equation (7. Which IV assumption is not satisfied when: a.fa + B^ Suppose that the assumptions in Key Concept 4.1.i.3 are satisfied.d. b.4 hJSLSXj)2 consistent? 457 Exercises \jL2. Zj = Wp.2.8 + «.] Would your answer change if R1 = 0. b. McNeil.05 and n .f l . How could you determine whether this instrument is relevant? How could you determine whether this instrument is exogenous? the purposes of this question suppose that the sample is very large and the TSLS estimators are essentially identical to fa and fa.05 and n = 100. a. a. and Newhouse (1994) used as an instrument the difference in distance to cardiac catheterization and regular hospitals. Wf-lforalH? d. Z.4 are satisfied.7") In an instrumental variable regression model with one regressor. Consider the instrumental variable regression model Yj = fa + faXj + faWj + Ui.)? b.2 of Equation (12. Construct a 95% confidence interval for the change in demand. Does this suggest that E(uj\Zy./-statistic is J =-18.

458

CHAPTER 1 2 Instrumental Variables Regression f 12.9 ) A researcher is interested in the effect of military service on human capi - — tal. He collects data from a random sample of 4000 workers aged 40 and runs the OLS regression Ys~ fa •+- B\Xt 4- «„ where Yt is the worker's annual earnings and X, is a binary variable that is equal to 1 if the person served in the military and is equal to 0 otherwise. a. Explain why the OLS estimates are likely to be unreliable. (Hint: Which variables are omitted from the regression? Are they correlated with military service?) b. During the Vietnam War there was a draft, where priority for the draft was determined by a national lottery. (Birthdates were randomly selected and ordered 1 through 365. Those with birthdates ordered first were drafted before those with birthdates ordered second, and so forth.) Explain how the lottery might be used as an instrument to estimate the effect of military service on earnings. (For more about this issue, see Joshua D. Angrist, "Lifetime Earnings and the Vietnam Era Draft Lottery: Evidence from Social Security Administration Records," American Economic Review. June 1990:313-336.) 12.10 Consider the instrumental variable regression model Yj = fa + ByX, -t- faW, + ttj, where Z;- is an instrument. Suppose that data on Wt are not available and the model is estimated omitting Wt from the regression. a. Suppose that Z, and W, are uncorrected. Is the IV estimator consistent9 b. Suppose that Z, and Wt are correlated. Is the IV estimator consistent? E12.2

Empirical Exercises

459

shipping price and other factors from 1880 to 1886.4 A detailed description of the data is contained in JECJDescription available on the Web site. Suppose that the demand curve for rail transport of grain is specified as ln(Qj) = fa + fa\n(I*) + falcej + 'L^fa^jSeasjj + «,-, where Qj is the total tonnage of grain shipped in week i, Pt is the price of shipping a ton of grain by rail, lcex is a binary variable that is equal to 1 if the Great Lakes are not navigable because of ice, and Seasj is a binary variable that captures seasonal variation in demand. Ice is included because grain could also be transported by ship when the Great Lakes were navigable. a. Estimate the demand equation by OLS. What is the estimated value of the demand elasticity and its standard error? b. Explain why the interaction of supply and demand could make the OLS estimator of the elasticity biased. c. Consider using the variable cartel as instrumental variable for ln(P). Use economic reasoning to argue whether cartel plausibly satisfies the two conditions for a valid instrument. d. Estimate the first-stage regression. Is cartel a weak instrument? e. Estimate the demand equation by instrumental variable regression. What is the estimated demand elasticity and its standard error? f. Does the evidence suggest that the cartel was charging the profitmaximizing monopoly price? Explain. (Hint: What should a monopolist do if the price elasticity is less than 1?) How does fertility affect labor supply? That is, how much does a woman's labor supply fall when she has an additional child? In this exercise you will estimate this effect using data for married women from the 1980 U.S. Census.5 The data are available on the textbook Web site www.pearsonhighered.com/ stock_watson in the file Fertilily and described in the file Fertility.Description. The data set contains information on married women aged 21-35 with two or more children. a. Regress weeksworked on the indicator variable morekids using OLS. On average, do women with more than two children work less than women with two children? How much less? These data were provided by Professor Robert Porter of Northwestern University and were used in his paper "A Study of Cartel Stability: The Joint Executive Committee, 1880 188ri," The BctlJournal of Economics, 148?, 14(2). 301-314. ' 5 These data were prov'ded by Professor William Evans of the University of Maryland and were used in his paper with Joshua Angrist, "Children and Their Parents' Labor Supply: Evidence from Exogenous Variation in Family Size." American Economic Review. 1*)98.88(3): 450-477.
4

Empirical Exercises E12.1 During the 1880s, a cartel known as the Joint Executive Committee (JEC) controlled the rail transport of grain from the Midwest to eastern cities in the United States. The cartel preceded the Sherman Antitrust Act of 1890, and it legally operated to increase the price of grain above what would have been the competitive price. From time to time, cheating by members of the cartel brought about a temporary collapse of the collusive price^setting agreement. In this exercise, you will use variations in supply associated with the cartel's collapses to estimate the elasticity of demand for rail transport of grain. On the textbook Web site www.pearsonhighered.com/stock_watson, you will find a data file JEC that contains weekly observations on the rail

460

CHAPTER 1 2 Instrumental Variables Regression b. Explain why the OLS regression estimated in (a) is inappropriate for estimating the causal effect of fertility (morekids) on labor supply (weeksworked). c. The data set contains the variable samesex, which is equal to 1 if the first two children are of the same sex (boy-boy or girl-girl) and equal to 0 otherwise. Are couples whose first two children are of the same sex more likely to have a third child? Is the effect large? Is it statistically significant? d. Explain why samesex is a valid instrument for the instrumental variable regression of weeksworked on morekids. e. Is samesex a weak instrument? f. Estimate the regression of weeksworked on morekids using samesex as an instrument. How large is the fertility effect on labor supply? g. Do the results change when you include the variables ageml, black, hispan, and othrace in the labor supply regression (treating these variable as exogenous)? Explain why or why not. E12.3 (This requires Appendix 12.5) On the textbook Web site www.pearsonhighered ,com/stock_watson you will find the data set Wcaklnstrument that contains 200 observations on (Yj. Xj, Z ( ) for the instrumental regression Y, = B{t+faXj + Uj. a. Construct faSLS, its standard error, and the usual 95% confidence interval for B\. b. Compute the /-"-statistic for the regression of Xt on Z,-. Is there evi dence of a "weak instrument'' problem? c. Compute a 95% confidence interval for B^ using the Anderson-Rubin procedure. (To implement the procedure, assume that - 5 < 0, < 5.) d. Comment on the differences in the confidence intervals in (a) and (c). Which is more reliable? 12.3 Large-Sample 1 2 . 2 Derivation of the

Large-Sample Distribution of the TSLS Estimator

461

adjusted) average retail cigarette price per pack during the fiscal year, including taxes. Income is real per capita income.The general sales tax is the average tax, in cents per pack. due to the broad-based slate sales tax applied to all consumption goods. The cigarette-specific tax is the tax applied to cigarettes only. All prices, income, and taxes used in the regressions in this chapter are deflated by the Consumer Price Index and thus are in constant (real) dollars. We are grateful to Professor Jonathan Gruber of MIT for providing us with these data.

Formula (12.4)

f o r t h e T S L S E s t i m a t o r in E q u a t i o n

The first stage of TSLS is to regress Xt on the instrument Z,- by OLS and then compute the OLS predicted value X„ and the second stage is to regress K, on Xj by OLS. Accordingly, the formula for the TSI.S estimator, expressed in terms of the predicted value X„ is the formula for the OI.S estimator m Key Concept 4.2. with X, replacing A,-.That is, f}fSLS =- s^y/s*, where sj is the sample variance of A'; and s^y is the sample covariance between Yj and Xj. Because AJis the predicted value of A", from the first-stage regression,^ = TTQ + K\Zi% the definitions of sample variances and covariances imply that s- — 7T\sZY and sj,~ TT\SZ (Exercise 12.4). Thus, the TSLS estimator can be written as faSLS = s^yfs^^ S^K^ST). Finally, TT1 is the OLS slope coefficient from the first stage of TSLS, so 77. = s7Xjs\. Substitution of this formula for 7 T 1 into the formula faSLS = sZY/(^T]Sz) yields the formula for the TSLS estimator in Equation (12.4).

Distribution of the T S L S 12.1 T h e Cigarette C o n s u m p t i o n Panel Data Set

Estimator

'The data set consists of annual data tor the 48 contiguous U.S. states from 1985 to 1995. Quantity consumed is measured by annual per capita cigarette sales in packs per fiscal year, as derived from state tax collection data. The price is the real (that is, inflation

This appendix studies the large-sample distribution of the TSLS estimator in the case considered in Section 12.1, that is, with a single instrument, a single included endogenous vari able, and no included exogenous variables. To start, we derive a formula for the TSLS estimator in terms of the errors that forms the basis for the remaining discussion, similar to the expression for the OLS estimator in

506

CHAPTER 13 Experiments and Quasi-Expenments small classes. Suppose that some principals succumbed to this pressure and transferred some children into the small classes. How would such transfers compromise the internal validity of the study? Suppose that you had data on the original random assignment of each student before the principal's intervention. How could you use this information to restore the internal validity of the study? 13.4 13.5 Explain whether experimental effects (like the Hawthorne effect) might be important in each of the experiments in the previous three questions. Section 12,1 gives a hypothetical example in which some schools were damaged by an earthquake. Explain why it is an example of a quasi-experiment. How could you use the induced changes in class sizes to estimate the effect of class size on test scores? -voJJ^^* * ( Average SAT score (X) Standard deviation of SAT score (sx) Number of men , Number of women Treatment Group 1241 93.2 55 45

Exercises

507

Control Group 1201 97.1 45 55 ^

a. Estimate the average treatment effect on test scores. b. Is there evidence of nonrandom assignment? Explain. /

l^v» o°"

K

n

,<>*\ A Exercises % \

o

13.4 ) Read the box"What Is the Effect on Employment of the Minimum Wage?" V -/ in Section 13.4. Suppose, for concreteness, that Card and Krueger collected their data in 1991 (before the change in the New Jersey minimum wage) and in 1993 (after the change in the New Jersey minimum wage). Consider Equation (13.7) with the W regressors excluded. a. What are the values of X,„ Gj. and D. for: i. A New Jersey restaurant in 1991? > , c ^ / ^
X l

Using the results in Table LK1, calculate the following for each grade: an estimate of the small class treatment effect, relative to the regular class; its standard error; and its 95% confidence interval. (For this exercise, ignore the results for regular classes with aides.) For the following calculations, use the results in column (4) of Table 13.2. Consider two classrooms, A and B, with identical values of the regressors in column (4) of Table 13.2, except that: a. Classroom A is a "small class" and classroom B is a "regular class." Construct a 95% confidence interval for the expected difference in average test scores. b. Classroom A has a teacher with 5 years of experience and classroom B has a teacher with 10 years of experience. Construct a 95% confidence interval for the expected difference in average test scores.

_ '

ii. A New Jersey restaurant in 1993? l ( . A - ' ' ^ ' ' ' * iii. A Pennsylvania restaurant in 1991? / ^ O I G W - O I ^ V iv. A Pennsylvania restaurant in 1993? v
r

^ 3.2 I3 J

0,&v*0v^

b. In terms of the coefficients fa, 8 L, fa, and y33, what is the expected number of employees in: i. A New Jersey restaurant in 1991? P " * ii. A New Jersey restaurant in 1993? in. A Pennsylvania restaurant in 1991? iv, A Pennsylvania restaurant m 1993? Po
A

i-)^U

\ r\(rrz

vO

* **
r5

-"

{-'&*^ i'°

:t;

c. Classroom A is a small class with a teacher with 5 years of experience and classroom B is a regular class with a teacher with 10 years of experience. Construct a 95% confidence interval for the expected dif>i ference in average test scores. (Hint: In STAR, the teachers were randomly assigned to the different types of classrooms.) ^> i> —r t ^i° >-o' l&y d. Why is the intercept missing from column (4)? Suppose that, in a randomized controlled experiment of the effect of an SAT preparatory course on SAT scores, the following results are reported:

;<V c^ K)> ' * ' tm-*' d. Explain why Card and Krueger used a differences-in-differences esti^ - V »H.,»L^O ^ vo jrjl malor of the causal effect instead of the "New Jersey after—New '* * At A ^ r f ^ ^ ^ '*** & / J e r s e y before" differences estimator or the " 1993 New Jersey - 1 9 9 3 iX vX ' j> (p. ^ e^p*-**^ Pennsylvania" differences estimator. *n C M J ' - - - »\ 13.5 Consider a study to evaluate the effect on college student grades of dorm room Internet connections. In a large dorm, half the rooms are randomly wired for high-speed Internet connections (the treatment group), and final

n

c. In terms of the coefficients fa, B\, fa, and 8\$, what is the average 1? causal effect of the minimum wage on employment?

(Hint: Use the homoskedasticity-only formulas for the variance of the OLS estima lor in Appendix 5. 13. The economics majors in the treatment group provide access to their Internet connection to those in the control group. Show that the OLS estimator of fi} is the differences-in-differences estimator in Equation (13. and let fif^1"'" ~diffs denote the differences-in-differences estimator. where (vh X„ fa. D. (Hint: Use the definition of the covariance and that.'Xfi = 0. Show that n\-ar(^ ') » (or + a n)/var(Xi2). Let ArJ= Y}2 Yn.T2* ** r c w j V c. so that Xit = 1 for t = 2 for the individuals in the treatment group and let Xj.1) is taken before the experiment and the second observation (t = 2) is for the post-treatment period. the OLS estimator in a regression of Ya on XJ2 with an intercept. a.fa *• BxXt + u{. based purely on efficiency considerations? 13. that is. E[BU\X^ = fa and E[v. Engineering students assigned to the control group put together a local area network so that they can share a private wireless Internet connection that they pay for jointly. random variables with fa = E(faj) Pi-E(faj). but ignore the Wregressor.1 V ° n over time. Suppose that outliers are rare so that (uh Xj) have finite fourth moments.„ a. that is.6) (in the case of a single W regressor. Consider the panel data regression model with fixed individual and time effects and individual characteristics Wj that do not change c. b. Let the treatment be binary. and c. (Hint: See Section 8. Bu) are i. n observations). Midway through the year all the male athletes move into a fraternity] and drop out of the study (their final grades are not observed). are individual fixed effects.) . + vit. and Wh and the a's and fas are unknown coefficients. where the first observation (/ .fference 2 2 where C.1. Let pffc™1** denote the differences estimator. Based on your answers to (a) and (b).-. = 1 if the individual is in the treatment group and G(.2)..3 are satisfied. is the product of D.6 Suppose that there are panel data for T = 2 time periods for a randomized controlled experiment. based on the OLS regression of &Yi ~ Yi2 .= 0 if the individual is in the control group. 8 / S u p p o s e that you have the same data as in Exercise 13. when would you prefer the differences-in-differences estimator over the differences estimator. cov(u (i . Which of the following pose threats to internal validity. = 0 otherwise. such as gender. X Wj) + faD.10). = CLJ + • faXj.d. is an error term.7 (panel data with two periods. 13.0. ^ d. where a. the estimator of 8. . ^ 1 if the i* individual is in the treatment group and t — 2..4).. a.3).10 Consider the regression model with heterogenous regression coefficients Yj = faj + fajXji Vj. and Xls — 0 otherwise. The art majors in the treatment group never learn how to access their Internet accounts. r^ ^ Exercises 509 K>^' wf*U . so r . Suppose that E[faj\Xj] . Show that assumptions 1 and 2 of Key Concept 4. Bu and Xj are independently distributed. X W. Further suppose that the treatment effect can be modeled using the specification Yj^aj + faX^+iij. Show that the model can be written as Y. D. Show that nvar(Miffs'in'dlfh) JO-Jfowhy?) > 2a2Jvar(Xl2). Consider the population regression model Yj. b.11)] with a mean of zero and a variance of a2 and «. ua) . where ^-(Boi-fa) + (Bu~B1)Xj+vi. Suppose that the treatment is binary. that is.0 for all /.9 Derive the final equality in Equation (13. (Hint: Note that .Yn against AA!) =. + fa(D.Xn and an intercept.1) from this population regression model. -^AJM ~s ^ ~* cK ^ U 3 . because the actual treatment Xt is random. Derive Equation (13.7 Suppose that you have panei data from an experiment with T = 2 periods (so t . suppose that Xj. where uh is homoskedastic.) 13.i. Is it appropriate to use OLS and the methods of Chapters 4 and 5 to estimate and carry out inference about the average values of B0i and faj? I . for a fee.fa.) b.Xi2 . Show that E[UJ\XJ] = 0. ' *.1.508 CHAPTER 1 3 Experiments and Quasi-Expenments course grades are collected for all residents. is the binary variable that equals 1 if / -= 2 and equals 0 if t = 1. where at are individual-specific effects [see Equation (13. d. and cov(u. and why? a. Consider the alternative regression model Yj^fa + faX^ + faGj + faDt + Uj.

A financial analyst claims to have a new model that makes better predictions than the random walk model. and to compare different forecasting models. ^ u . Show that E(Yt) = fa/(l .123 Y.^). 8.377 + 0. This problem uses data on this index for the United Stales.056) . Key Terms first lag (520) / h lag (520) first difference (520) autocorrelation (522) serial correlation (522) autocorrelation coefficient (522) / h autocovariance (522) autoregression (526) forecast error (527) root mean squared forecast error (RMSFE) (528) /?"' order autoregressive model [AR(p)] (529) autoregressive distributed lag (ADL) model (534) ADL(p. Suppose that you suspected that the intercept in Equation (14. Is this correct? Why? b.068) (0.055) (0. — fa + fa Y. 7./IP.-2 + 0. and. to estimate the root mean squared forecast error. A series that contains a stochastic trend is nonstationary.001 V.062) (0. 14. (Hint: Read Key Concept 14. Let Yt -= 1200 X \n(IP. potentially leading to biased estimators. . It implies that the percentage changes in stock prices are unforecastable. shows the monthly percentage change in IP. Many financial economists believe that the random walk model is a good description of the logarithm of stock prices.2 The index of industrial production (IP. Suppose that you calculated the first difference of this series. and misleading inferences. Suppose that a forecaster estimates the following AR(4) model for Y.02.4 Exercises 14. The OLS estimator and /-statistic for the coefficient of a regressor with a stochastic trend can have a nonstandard distribution.078) (0. then OLS estimates neglecting this instability are unreliable for statistical inference or forecasting._4. All regressions are estimated over the sample period 1960:1 to 2000:12 (that is. January 1960 through December 2000). A random walk stochastic trend can be eliminated by using first differences of the series.) b. a.068Yt„3 + 0.3 14. violating the second least squares assumption in Key Concept 14. Show that E(Yt) = E(Y. The A D F statistic can be used to test for a stochastic trend.q) (534) stationarity (535) weak dependence (537) Granger causality statistic (538) Granger causality test (538) forecast interval (540) Bayes information criterion (BIC) (543) Akaike information criterion (AIC) (543) trend (546) deterministic trend (546) stochastic trend (546) random walk (547) random walk with drift (548) unit root (549) spurious regression (550) Dickey-Fuller test (551) Dickey-Fuller statistic (552) augmented Dickey-Fuller (ADF) statistic (552) break date (557) Quandt likelihood ratio (QLR) statistic (558) pseudo out-of-sample forecasting (561) Review the Concepts 14. if a discrete break is found. Does a 95% confidence interval include 8^ = 1? Explain.: Yt= 1.6. The QLR statistic can be used to test for a break.1 Exercises 571 Look at the plot of the logarithm of G DP for Japan in Figure 14.! + 0.) is a monthly time series that measures the quantity of industrial commodities produced in a given month. the regression function can be re-estimated in a way that allows for the break.1 Consider the AR(1) model Y.B{). If the population regression function changes over time. The forecaster states that Y. Explain how you would examine the analyst's claim that his model is superior. A researcher estimates an AR(1) with an intercept and finds that the OLS estimate of fa is 0. measured in percentage points per annum. Does this time series appear to be stationary? Explain.318^.2c.^). How would you modify the equation to incorporate this change? How would you test for a change in the intercept? How would you test for a change in the intercept if you did not know the date of the change? 14. with a standard error of 0. a.95. Would it appear to be stationary? Explain. Pseudo out-of-sample forecasts can be used to assess model stability toward the end of the sample. (0.2 14.570 CHAPTER 1 4 Introduction to Time Series Regression and Forecasting 6. Suppose that the process is stationary.17) changed in 1992:1.5. inefficient forecasts.

and forecast errors: a. Show that the conditional mean squared forecast error is minimized when f. * 0.6 over the same sample period. Let fi-i denote some forecast of Y„ where the subscript t .650 2000:10 148.2 augments her AR(4) model for IP growth to include four lagged values of AR. Use the A D F statistic to test for a stochastic trend (unit root) in ln(/P). Do interest rates help to predict IP growth? Explain.c)2. Call these variables e. Let E[(Yt — ft-^Y. .061 • + . artificial data are generated using a computer. Worried that she might have included too few or too many lags in the model..0.007) (0.175 2 2S.-v2 is .. This makes it possible to compute the distribution of statistics for known models when mathematical expressions for those distributions are complicated (as they are here) or even unknown.595 2000:9 L48. a researcher tests for a stochastic trend in ln(/P.075) (0.4 Exercises 573 2000:8 148. In this exercise. Treasury bills (measured in percentage points at an annual rate). 4. c. e100. Set Yi .333Aln(//> . Show that cov(«f.1 on ft ^ indicates that the forecast is a function of data through date t -1.2 ). 3. The sum of squared residuals from each of these estimated models is shown in the table. The researcher also regresses AR. The resulting Granger-causality F-statistic on the four lags of IP growth is 2. Use the BIC to estimate the number of lags that should be included in the autoregression.S. . { = fy t.) t. 100.6 14. Worried about a potential break. Does IP growth help to predict interest rates? Explain.c)2} = o-^ + (p. ] be the conditional mean squared error of the forecast ft-\.) using the following regression: &HTP. on a constant.where R.162Aln(/^ 2 ) (0. Yt. Yt and XT. b.2.w .00004r . e2 . Let u.i.317 14. Consider the problem of forecasting Y. where Y^x = E(Y^Y.Y. conditional on Y observed through date t -1. is 2. b.d.538 3 28. are independently distributed random walks. .2? Explain. . a. for / = 2. Use your computer to generate a sequence of T= 100 i.45.572 CHAPTER 1 4 Introduction to Time Series Regression and Forecasting Use this AR(4) to forecast the value of Y. a. [Hint: Use Equation (2. using data on ^ _ b Yt.0.14). u.3"?8 6 28.5 Prove the following results about conditional means.00001) (0. Worried about potential seasonal fluctuations in production. The resulting QLR statistic was 3. e.6. and four lags of IP growth. b.055) where the standard errors shown in parentheses are computed using the homoskedasticity-only formula and the regressor "/" is a linear time trend.The estimated coefficient on Y. four lags of AR. forecasts. j) = 0 for. the forecaster estimates AR(p) models for p .660 2000:11 148. The Granger-causality F-statistic on the four lags of AR.. Is there evidence of a break? Explain.. Using the same data as in Exercise 14.^.300 The forecaster in Exercise 14.) = 0.27). you will generate data so that two series.206 2000:12 147. J + 0. .35. 12 to the autoregression. Do these results support the specification used in Exercise 14.024) (0. Yt_2. she computes a QLR test (with 15% trimming) on the constant and AR coefficients in the AR(4) model.] In this exercise you will conduct a Monte Carlo experiment that studies the phenomenon of spurious regression discussed in Section 14.ej and Y..973 14.391 5 28.1 .2.054 with a standard error of 0.0. and then these artificial data are used to calculate the statistics being studied.3 1 29. Show that E[(W . Let W be a random variable with mean nw and variance tj\ and let c be a constant.87.e.0. Do the results differ if you use the AIC? AR Order SSR 14. (Hint: Review Exercise 2. is the interest rate on threemonth U. the forecaster adds Y.27. standard normal random variables.393 4 28. In a Monte Carlo study. in January 2001 using the following values of IP for August 2000 through December 2000: Date IP 2000:7 147.0181n(/Ff .. .053.\.) c. denote the error in Equation (14. Is this coefficient statistically significant? d. The specific steps are as follows: i.

14. b. p.. 50%.1.ax and X. is a binary variable Empirical Exercises On the textbook Web site www. Suppose that YT = 102.. and there are q = 5 restrictions. a.. the logarithm of real GDP. iii.3. . onto a constant and X. «100.. for example. a. .i. Yt has a pronounced seasonal pattern. T . l + *2).96. d. Answer the following questions using the values in Table 14. Derive the AR(2) coefficients for Y.. Suppose that q = 1.d. Suppose that Yt is the monthly value of the number of new home construction projects started in the United States. the data are described in the file USMacro_Description. = fa + faFebt + faMar. is i. Should the researcher reject the null hypothesis at the 5% level? 14. for example. and so forth. Because of the weather.pearsonhighered. d.b\ + c.1 . Show that Y. follows the stationary AR(1) model Y.2.. a2. Should the researcher reject the null hypothesis at the 5% level? b. Compute the first two autocovariances of Yv (Hint: Read Appendix 14. standard normal random variables.bqe...6 ("Critical Values of the QLR Statistic with 15% Trimming") and Appendix Table 4 ("Critical Values of the Fmpo Distribution").-q.8 Yt-X.. follows the AR(1) model AY.Feb. does the fraction of times that you reject the null hypothesis approach 5%.5 0.E(YT+l\Yr.2.).. for t ~~ 2. of T = 100 i. Set X\ -. and AY. What is the R2 of your regression? b. Use this algorithm to answer the following questions: a..574 CHAPTER 1 4 Introduction to Time Series Regression and Forecasting ii. where e. Derive the autocovariances for Y. ..com/stock_watson.71J-] where u..i + b2et-2 +•••-«. Use the r-statistic from (iii) to test the null hypothesis that Bx = 0 using the usual 5% critical value of 1. you will find a data file USMacro_Quarterly that contains quarterly data on several macroeconomic series for the United States.~ fa + faAYt a.2. Show that fa + fa = p.) c. follows an AR(2) model. (Hint: See Exercise 14. Show that the variance of Y. Let pJan denote the average value of housing starts in January and ix.AJ-j + a. and 95% percentiles of the distributions of the R2 and the f-statistic? In what fraction of your 1000 simulated data sets does the f-statistic exceed 1. with E(u. Compute the mean and variance of Ys. the quarterly growth rate of GDR In Empirical Exercises 14.Mar. The QLR F-statistic is 2.• • • + BuDec. and so forth.7 Suppose that Y. Compute the first two autocorrelations of Y. 14. Repeat (a) 1000 times. Compute the OLS estimator. Show that the values of fijan. a. saving each value of R and the f-statistic. pDec can be estimated from the OLS regression Y. = \n(GDPt).5. is a serially uncorrected random variable with mean 0 and variance o-. Regress Y. and the (homoskedastic-only) f-statistic testing the null hypothesis that Bx (the coefficient on Xt) is zero.Vto be independently distributed? Does this fraction seem to approachisome other limit as 7 gets large? What is that limit? 14.3 100. is var( Yt) — a2e( 1 -^ b\ +. Repeat (b) for different numbers of observations. /x/r^. Mar. The QLR F-statistic is 4. housing starts are low in January and high in June..11 Suppose that AY.d. Compute y r+ ij T ~.96 in absolute value? c. + ut..50 and T — 200. -J. 2 Empirical Exercises 575 equal to 1 if t is March. Show that E(%) = fa. + u„ where Feb. b.. Compute Y. What are the 5%.9 The moving average model of order q has the form Yt = fa + e.1. 14. Use your computer to generate a new sequence.10 A researcher carries out a QLR test using 25% trimming. is a binary variable equal to 1 if t is February. Should the researcher reject the null hypothesis at the 5% level? c. The QLR /-"-statistic is 3. the regression R2.Man — V-Dee denote the average values in the other months. as a function of fa and fa.) b. as it should because you have generated Vand . As the sample size increases.) = 0 and var(wr) = 9. Show that p. Construct a histogram of the R2 and r-statistic.= 0 for/ > q.i. + V . ax. Run the algorithm (i) through (iii) once.

Repeat the calculations reported in Table 14.1). is stationary around a deterministic trend. Express your answer in percentage points at an annua] rate. is the value of the index at the daily close of the New York Stock Exchange. Consumer Price Index is measured using monthly surveys and is compiled by the Bureau of Labor Statistics (BLS). Construct pseudo out-of-sample forecasts using the A DL( 1.What is the estimated AR(1) coeffi cient? [s the coefficient statistically significandy different from zero? Construct a 95% confidence interval for the population AR(1) coefficient. go to the course Web site.4) model to the AR(1) model.4 E14. Is the variable \n(dividend yield) highly persistent? Explain. b. What are the units of the autocorrelations (quarterly rates of growth. use the sample period 1955:1-2009:4 (where data before 1955 may be used.4) model for A Y. Construct pseudo out-of-sample forecasts of excess returns over the 1983:1-2002:12 period using regressions that begin in 1932:1. e.4) model. the data are in the file Stoek„Returns_1931_2002 and are described in the file Stock_Returns_1931^2002_Description. denote the interest rale for three-month treasury bills.6 a. Estimate the first four autocorrelations of A Y.. as initial values for lags in regressions). as necessary.+ A ^ + AY. Estimate an AR(1) model for A^. as additional predictors. Next.. Test for a break in the coefficients on the constant term and coefficients on the lagged values of AR using a QLR test.3 using regressions esti mated over the 1932:1-2002:12 sample period. Is there evidence of a break? E14.-2 + A^_ 3 )/4. because . The federal funds rate data are the monthly average of daily rates as reported by the Federal Reserve. c. where you will find an extended version of the data set described in the boxes.1 E14.6. (That is.7 using regressions estimated over the 1932:1-2002:12 sample period. both are for the final month in the quarter.3 E14. or no units at all)? E14. E14. Are any of the forecasts biased? Which model has the smallest root mean squared forecast error (RMSFE)? How large is the RMSFE (expressed in percentage points at an annual rate) for the best model? Read the boxes "Can You Beat the Market? Part I" and "Can You Beat the Market? Part II" in this chapter. Estimate AR(3) and AR(4) models. a. [Hint: Multiply the sample mean in (a) by 400.] c. where NYSE. Estimate an ADL(1. Is the AR(2) coefficient statistically significantly different from zero? Is this model preferred to the AR(1) model? c. use BIC to choose the number of lags in the AR model. Is the Granger causality F-statistic significant? c.7 Time Series Data Used in Chapter 14 577 c. by how much has the R2 changed? b. llPBENDIX? 14.1 a.S. The daily percentage change in the NYSE Composite Index was computed as 100Aln(/VyS£r). Estimate an AR(2) model for AY.. The U.2 a.. a. Express the mean growth rate in percentage points at an annual rate.) b. using lags of AR. and the dollar/ pound exchange rate data are the monthly average of daily rates. Comparing the ADL(1. and so forth. Repeat the calculations reported in Table 14. (//') How many lags does A1C choose? i Use an augmented Dickey-Fuller statistic to test for a unit autogressive root in the AR model for Yr As an alternative. Do the results in (a) through (d) suggest any important changes to the conclusions reached in the boxes? Explain. Compute the pseudo out-of-sample forecast errors for each model. using a QLR test. d.5 T i m e S e r i e s D a t a U s e d in C h a p t e r 14 Macroeconomic time series data for the United States are collected and published by various government agencies. compute AY]q9Q:i\m9:4> A ^ y ^ j w a j . suppose that Y. b. Estimate the standard deviation of A Y.576 CHAPTER 1 4 Introduction to Time Series Regression and Forecasting through 14. percentage points at an annual rate. Japanese GDP data were obtained from the OECD. d. (/) Using the estimated AR(1) through AR(4) models. E14. Construct pseudo out-of-sample forecasts using the AR(1) model beginning in 1989:4 and going through the end of the sample. d. Let R. b. Construct pseudo out-of-sample using the following "naive" model: AYl+l/t~-(AY.The unemployment rate is computed from the BLS's Current Population Survey (see Appendix 3. Estimate the mean of AY. Test for a break in the AR(1) model for A Y.The quarterly data used here were computed by averaging the monthly values.

+ ••• +AX. present. Show that corr(«„ u.i. Show that cov(«„ u. Derive Equation (15. b.1) applied to the quasidifferenced regression Equation (15. 1 2 ?%-= and var(X) = -— e -^. c. + u„ where u. [Hint: Rearrange the formula in (b).^p+x^Xt. i. Show that a2. i.1.<pxu.^x + u. GDP growth is observed. with u. Suppose that X. is strictly exogenous.) . oil prices are set randomly and the central bank holds interest rates constant and in particular does not respond to the oil price changes. .25.^ + 2. In experiment B. is independent of u-} for all t and.+x. . In both. 8 ^ + u„ where X. (Hint: Note that X.lS2\YtfaY^i). b.d.1.iyl?. Sx = Bu 82 = (3X± fa./3o small when T is large.ctf 1 " 7i b.To which experiment.B{ •*-fa+ fa (etc.| < 1. Explain why the GLS estimator of (3X is not consistent. follows the stationary AR(1) model ut — (j>xu.23).7 625 Consider the regression model Y. with mean 0 and variance rj\ and |4>. and fr in Equation (15. . c.1 correspond? Suppose that oil prices are strictly exogenous.3 Consider two different randomized experiments.14). i. Exercises ii. Are the coefficients significantly different from zero? 15.0^..p. = yxXt i + e. exogenous? Is X.d. and e. is independent of u. 01" 15. i. Derive an expression for /co15. Show that the OLS estimator is fa = T 2 . strictly exogenous (past. 1 .0 . Is the OLS estimator of fix consistent? Explain. " — 17 * a. = <pxu. Show that the infeasible GLS estimator B GLS <P1 1 + 4% [Hint: Use the omitted variable formula (6. Show that the (infeasible) GLS estimator is j8f LS = (1 . d. with mean 0 and varianceCT~and \(px\ < 1. — 25 and 0. Show that var(«.A or B.4>i) '\T.-j) = cp[ and corr(Jtf„ X. — fa + BXX. oil prices are set randomly and the central bank reacts according to its usual policy rules in response to economic conditions. Construct 95% confidence intervals for your answers to (a). The HAC /-"-statistic testing whether the coefficients on 0 . + * * . with u.'. the regressor X. Show that (3§LS can be written as p§LS = (T-i) • • ^ S l J + d . does the dynamic causal effect estimated in Exercise 15.^) — y(.) and cov(^.8 Consider the model in Exercise 15. Derive the impact effect of X on Y.-X + u.i. = o-xcrl.).y(vav(Xt). c. . a. Derive the first five cumulative multipliers. Suppose that oil prices jump 25% above their previous peak value and stay at this new higher level (so that O. exogenous (past and present)? Is X. with e.7 with X. a.BXX. What is the predicted change in interest rates for each quarter over the next 2 years? b. + 1 = Ol+2 = • • • = 0). for all t and i. In experiment A. with mean 0 and variance o-~ and \(f>x < 1. and future)? b.d. where ax is the variance of X and cr2u is the variance of u.4 15. follows the stationary AR(1) model X. Discuss how you could improve on the estimates of the dynamic multipliers in Exercise 15.10 Consider the A D L model Y. X.4) and show that S0 = B0. What is the effect of this change in oil prices on the level of interest rates in period t f 8? How is your answer related to the cumulative multiplier? d. . [Hint:The GLS estimator of /3Q is (1 .] d. = u.^ r 1 ^ lrl(YT~<pxY).+x. and its lags are zero is 4. = i ^. a. b. where u. x f u. = AX.9 15.u.--3A+ OAY. — B0 4. Derive the first five dynamic multipliers. Is X. strictly exogenous? 15.i. + w„ where u. with u.d.) Consider the regression model Y. Consider the terms o-2. follows the stationary AR(1) model ut . . Now suppose that oil prices are exogenous in the regression in Exercise 15. Is X. •) = rf>ivar(«. = fa + u. including changes in the oil price.i. c. JDerive the difference BQ . follows the stationary AR(1) model u. Suppose that X.) = 15.7) from Equation (15. d. with mean 0 and variance v2 and |yi | < 1.-.5 15. Why?] c.cf>x )~* multiplied by the OLS estimator of a0 in Equation (15. 5 3 .23)]. Derive the long-run cumulative dynamic multiplier. i. and discuss why it is likely to be a.624 CHAPTER 1 5 Estimation of Dynamic Causal Effects a.6 Consider the "constant-term-only" regression model Y.

statistically significantly different from zero? e. These series are alternative measures of consumer prices in the United States. negative? c. obtained from the National Oceanic and Atmospheric Administration (NOAA) of the U. Is there statistically significant evidence that the mean inflation rate for the CPI is greater than the rate for the PCED? g. Construct graphs like those in Figure 15. CPIa n d V f c t u .pearsonhighered. is serially correlated? Explain.S.626 CHAPTER 1 5 Estimation of Dynamic Causal Effects Empirical E x e r c i s e s E15. What are the mean and standard deviation of ip _growth over the 1952:1-2009:12 sample period? b. The freezing degree days series was constructed from daily minimum temperatures recorded at Orlando-area airports.1. The FDD series was constructed so that its timing and the timing of the orange juice price data were approximately aligned. f.1200 X in[CPI(t)lCPI(t . Let Trf 7 . for February is the percentage change in real orange juice prices from mid-January to mid-February. is the difference. Estimate a distributed lag model of ip^growth onto current and 18 lagged values of 0.The orange juice price series was divided by the overall PPI for finished goods to adjust for general price inflation.2 In the data file USMacro Monthly. What value of the HAC standard truncation parameter m did you choose? Why? f.. Compute the sample mean of Yt.100 X \n(IPJIP. Accordingly. although the exact date varies from month to month.. that is. and 95% confidence intervals. E15. and Y. summed over all days from the 11th to the 10th. Consider the "constant-term-only" regression: Y. -rrfc ED is the monthly rate of price inflation 15.com/stock„watson in the file USMacro_Monthly. Comment on the real-world size of the multipliers.tr™ . equal to zero? Why aren't some values of O.Thus %ChgP. Are these point estia. Show that the population mean of Y is equal to the difference of the population means of the two inflation rates. FDD is the maximum of zero and 32 minus the minimum daily temperature. . cumulative multipliers. a. . .2 showing the estimated dynamic multipliers. Compute the monthly growth rate in IP expressed in percentage points. e. carry out the following exercises. TT^CED = 1200 X ln[PCED(t)lPCED(t .1)]. Compute the sample means of TT^' mates consistent with the presence of economically significant substitution bias in the CPI? b. c. The orange juice price data are the frozen orange juice component of processed foods and feeds group of the Producer Price Index (PPI).1 T h e O r a n g e Juice D a t a Set The Orange Juice Data Set 627 from the PCED. Plot the value of Or Why are so many values of O. Is there evidence of instability in 8Q1 Carry out a QLR test. then average CPI inflation should be systematically higher than PCED inflation. ). What value of the HAC standard truncation parameter m did you choose? Why? d. Taken as a group. the FDD series was constructed to be the number of freezing degree days from the 11th of one month to the 10tl1 of the next month. you will find data on two aggregate price series for the United States: the Consumer Price Index (CPI) and the Personal Consumption Expenditures Deflator (PCED). Specifically.c/>/ is the monthly rate of price inflation (measured in percentage points at an annual rate) based on the CPI. collected by the U. Explain why it is numerically equal to the difference in the means computed in (a). ip_growth. Department of Commerce. Construct a 95 % confidence interval for fa. Do you think that u. Using data from 1970:1 through 2009:12. for February is the number of freezing degree days from January 11 to February 10. d. Is Ot exogenous? Are the estimated multipliers shown in the graphs in (e) reliable? Explain.1 In this exercise you will estimate the effect of oil prices on macroeconomic activity using monthly data on the Index of Industrial Production (IP) and the monthly measure of O.1)]. Show that fa = E(Y). The CPI prices a basket of goods whose composition is updated every 5-10 years. and Y. Bureau of Labor Statistics (BLS series wpu02420301). so u. If this substitution bias is important. The data can be found on the textbook Web site www. the frozen orange juice price data are collected by surveying a sample of producers in the middle of every month. The PCED uses chain-weighting to price a basket of goods whose composition changes from month to month. .S. described in Exercise 15.. are the coefficients on 0 . Economists have argued that the CPI will overstate inflation because it does not take into account the substitution that occurs when relative prices change. Suppose that high demand in the United States (evidenced by large values of ip^growth) leads to increases in oil prices. and FDD.n f ™ . = fa + u.

^.0 + 0.0 + 1. e. Y. Verify Equation (16.).. + vx. [Hint: Analyze the numerator of 8 using analysis like that in Equation (16.-p f «„ where E(ut\Yt x. = ]n(GDP._ i A>.2 In this exercise you will compute pseudo out-of-sample two-quarter-ahead forecasts for AY beginning in 1989:4 through the end of the sample._ 1 . Derive the vector error correction model [Equations (16. what isPr(-3 < « . a. -t+ ••• + BpY. R.) b.O.+2i. Suppose that the distribution of ut conditional on lagged values of u. a2 ~ 1.i.21). (Hint: Use the law of iterated expectations E(u2) = E[E(u2\ut-X)\) b. M. a. . .10 Consider the cointegrated model Y.7 Xcointegrated? Xcointegrated? ii.0. 3)/4. . Show that Sf=ia.+x and error is the regression error.? b. What are the values of 6. J v . and X. 4-».0.. and consider the regression Yt = BX.25)..= (AY.+ S ^ . Empirical Exercises These exercises are based on data series in the data files USMacro_Quarterly and USMacro_Monthly described in the Empirical Exercises in Chapters 14 and 15. • • •) follows the A R C H ( l ) model oj = a0 + axu2-X. (That is. Construct iterated two-quarter-ahead pseudo out-of-sample forecasts using a VAR(4) model for A Y and A R. c. is 1(1). s 3 ) ? 16._2.< 1 for a stationary ARCH(p) model.0 for all t and /. -2 and derive values for o". where u f is i.....I). Are F a n d iii. Rearrange the regression so that it has the form shown in Equation (16.22) and (16. is /(l) and u{ is 1(0). . Are Yand 16.) a.20). = Xt.~ u.(Hint: Use the law of iterated expectations E(u2) = E[E(u2\u.! -r 0.) = <V(1 ~ a\). Show that (3 > \(x\ .var(w. w h a t i s P r ( . and 82 in terms of the coefficients in the VAR. 2 . [Hint: Use 2 ^ * 7 = ^LM . you will compute A Y^oa/i989.1) model. is 1(1) and u.1 Using quarterly data from 1955:1 through 2009:4.. where X. &\. b. follows the AR(p) model Y.3 < u.^2...) = 2. is stationary. yields Yt . Show that var(w. and future values of X.. b.6 16.) .4> A ^. follows the ARCH process. Let Y.. 2 . Y. ) = 0. In light of your answer to (a). that var(« f |u. isN(0. Suppose that X.. is 7(1). past. /V(0. a.x+& + ••• +/3 / . estimate a VAR(4) (a VAR with four lags) for A Y( and AR.0 .] 16.4 Suppose that Y. Show that ax + <px < 1 for a stationary GARCH(1.. Show that the iterated two-period-ahead forecast for Y can be written as Ytn-2 ~ 8xYt 2 + 82X.990:3/1990:1 • a n ^ s o forth.9 667 a../ J + / l | . do iterated multiperiod forecasts differ from direct multiperiod forecasts? Explain. Show that Yt+tif=fa\ faYt.23)] for X and Y. Extend the result in (a) to the ARCH(p) model. denote the 3-month Treasury bill rate. is 7(0) and u.lf u.-X)].. c. x.Let Y. ! ~ 2.3 Suppose that u.8 Consider the following two-variable VAR model with one lag and no intercept: X^faiYt 1+mAi^+w. ! . are mean zero serially uncorrelated random variables with E(vXtv2]) . and that the process for u.2*.8*. . Extend the result in (a) to the GARCH(1. + A^) 2 to show that 2 ^UY t = 2 ^ I V? l + 2S^ 1 y.666 CHAPTER 1 6 Additional Topics in Time Series Regression 16.7A. = £(*. v Xt = PuYt i + y 2 i*. E16.5 16.1) model. Which model has the smallest root mean squared forecast error? a. and 5. . = 6X. Suppose that E(ut\ut-X) ut. Should the VAR include more than four lags? E16.A Y.8 . Are Y and X cointegrated? Suppose that AY. Does AR Granger-cause A V? Does A Y Granger-cause A/?? b.1). Let E(u ) . Analyze the denominator using the law of large numbers. d. < 3)?Ifu. + u2t.-X.) be the unconditional variance of uv Show that var(». . x 2 ""V Empirical Exercises 16. + error.+AY.for/r>p. i. A y ? and solve for S M ^ A Y .3. 2 F AY. Construct iterated two-quarter-ahead pseudo out-of-sample forecasts using an AR(1) model. respectively. and •n-p^and 7rf C£ denote the inflation rates from the CPI and Personal Consumption Expenditures (PCE) Deflator.x + v2„ where vX( and v2. d.-X + A ^ . . Construct iterated two-quarter-ahead pseudo out-of-sample forecasts using the naive forecast AY. Suppose that X.cx2). 16. ] A regression of Y J onto current.5u2.).d.^ fa+ B^Y. Suppose that X.Ul\.x.

but you construct a 95% confidence interval for 8X using the homoskedasticity-only standard error formula in a large sample.1 . conditional on the regressors. What is the asymptotic distribution of AnBnl Use this asymptotic distribution to compute an approximate value of ?r(AnBn < 2). (693) bivariate normal p.25)]. Yj = (3xXt + Uj (so the true value of the intercept. (694) .3 17.B-Xj + «. jS^. and normally distributed.1 Exercises 689 Suppose that Assumption #4 in Key Concept 17. .4 Exercises 17.1. the researcher in the previous problem decides to compute the OLS estimator using only the observations for which x < 10. Derive the least squares estimator of 8X for the restricted regression model Yj . combined with the consistency of heteroskedasticity-robust standard errors. . For this reason. This problem is difficult enough with a single regressor. In contrast.f. if the first three least squares assumptions in Key Concept 17. but in applications with multiple regressors it is even more difficult to know the functional form of the conditional variance. or its estimate. .1 hold. which usually is a tall order. under Assumptions #1 and #2 of Key Concept 17. If the regression errors are i.Xi) is rarely if ever known.d. Although the WLS estimator is asymptotically more efficient than OLS. then using only the observations for which x > 10.1. then the heteroskedasticity-robust f-statistic has an asymptotic standard normal distribution under the null hypothesis. Is this estimator more efficient than WLS? 17.f. 17. n. where 0 ^ X < 20. a. implies that.i. Yj). b. Derive the asymptotic distribution of faLS under Assumptions # 1 through #3 of Key Concept 17. Show that faLS is linear [Equation (5. . The weighted least squares (WLS) estimator is OLS applied to a weighted regression. 3. but you construct a 95 % confidence interval for 8 x using the heteroskedastic-robust standard error in a large sample. For these reasons.24)] and. and the resulting inferences are reliable under very general conditions. Suppose that V a n d X a r e related by the regression Y— 1.0 + 2. conditional on the regressors.This is called the restricted least squares estimator (fixLS) of (3X because it is estimated under a restriction. it is our opinion that.1 is false. Suppose that Bn is a random variable that converges in distribution to a standard normal.1 is true. Draw a hypothetical scatterplot of the observations (Xj. var(ujiAJ). despite the theoretical appeal of WLS. Does WLS put more weight on observations with x < 10 or x > 10? Why? Instead of using WLS. and then average the two OLS of estimators. which poses a problem for using WLS in real-world applications. i* Key T e r m s convergence in probability (673) consistent estimator (673) convergence in distribution (675) asymptotic distribution (675) Slutsky's theorem (676) continuous mapping theorem (676) weighted least squares (WLS) (683) WLS estimator (684) infeasible WLS (684) feasible WLS (685) normal p. c. the homoskedasticity-only r-statistic has an exact Student tn_2 sampling distribution under the null hypothesis. which in this case is fa = 0. in modern statistical packages it is simple to use heteroskedasticity-robust standard errors. The asymptotic normality of the OLS estimator. is zero). A researcher has observations on Y and X. in particular.d. In addition. conditionally unbiased [Equation (5.d.688 CHAPTER 1 7 The Theory of Linear Regression with One Regressor In practice.i. Would this confidence interval be valid asymptotically? Suppose that An is a random variable that converges in probability to 3. heteroskedasticity-robust standard errors provide a better way to handle potential heteroskedasticity in most applications. where the conditional variance is var(uj\Xj = x) = 1 for 0 -£ x < 10 and var(Uj\Xj = x) 16 for 10 =s x < 20.2 Summary 1. where all variables are weighted by the square root of the inverse of the conditional variance. then fa has an exact normal sampling distribution. practical use of WLS confronts imposing challenges. Would this confidence interval be valid asymptotically in the sense that it contained the true value of B x in 95 % of all repeated samples for large nl Suppose instead that Assumption #4 in Key Concept 17.QX + U. 2. to implement WLS you must know the functional form of the conditional variance function. heteroskedasticity-robust standard errors can be used without needing to specify a functional form for the conditional variance. the functional form of var(uj.-.1 Consider the regression model without an intercept term. Review the Concepts 17.

c.2 Suppose that (A(-.12).)f(uj.-) = 0O + 6x\Xj\.Xn) > var^f-^.. where \Xj\ is the absolute value of Xj.3 17..16) under Assumptions #1 and #2 of Key Concept 17.V satisfy the central limit theorem? d. Show that E(u}\Xx Xn) = E(UJ\XJ).24). fa.f.. (Hint: Use Slutsky's theorem. #2. Exercises 17. Use the central limit theorem. where a2 is a constant. Xj) for / ^ /.) of (u. g.^).i. the mean squared error of 0 tends to zero). Derive the exact sampling distribution of the WLS estimator (treating 0O and 0X as known) of Bx. Suppose that W is a random variable with E(W4) < oo. Xj) is the joint p..5 17. Compare the conditional variance of fa in (d) to the conditional variance of the OLS estimator Bx (from the regression including an intercept) under the Gauss-Markov conditions..are i. Suppose that X and u are continuous random variables and (Xj. Att.4 Show the following results: 691 a... (Hint: Use the strategy of Appendix 3.3 and the Cauchy-Schwarz inequality.i.6 17. Which estimator is more efficient? Use the formulas for the variances to explain why. d.38). c Use the Cauchy-Schwarz inequality and the third least squares assumption in Key Concept 17.) This exercise fills in the details of the derivation of the asymptotic distribution of fa given in Appendix 4.Xn) f. i = 1 . 17. Suppose that 0O and 0X are known. Apply the central limit theorem and Slutsky's theorem to obtain the result in Equation (17..) b.. Is the OLS estimator of fa BLUE? Jnl£ > v i X MAO-V/HS^ n{fa ~fa) = b.690 CHAPTER 1 7 The Theory of Linear Regression with One Regressor d... Show that E(ujUpXx. . a. Xj) can be written as f(u. have eight moments. then B —*-• 0. where 6 might be biased. Show that the joint probability density function (p-d. a.8 Consider the regression model in Key Concept 17. a.3./'S"=iAJ. where f(u.11 Suppose that A and Y are distributed bivariate normal with density given in Equation (17. . that is.19) to derive the expression 17. = E(UJ\XJ)E(UJ\XJ) for / * j. Prove that the sample covariance is a consistent estimator of the population covariance. .9 Prove Equation (17.1 plus the assumption that A. and #5 hold.. ut). Does the term V ^ H V / O . a2). [Hint: Use Equation (17.X2. where v..d.7 p —> 1. #3.. Derive the conditional variance of BXLS under the Gauss-Markov conditions (Assumptions #1 through #4 of Key Concept 17. b. conditional on Xx. Xn.(A.d.1). with finite fourth moments. and «.n. of iij and X..var(^f LS |A' 1 Xn) under the Gauss-Markov conditions and use this expression to show that var(fa\Xx. Now consider the estimator fa = 2"=i>.1 and suppose that Assumptions #1.|A.1. 17. show ihat\fE(fo\Xx. Show that if fa is conditionally unbiased then it is unbiased.. Use Equation (17. Show that s2Jcr2 -•••£-> 1 implies that sufcru 17.10 Let 9 be an estimator of the parameter 6. Derive the exact sampling distribution of BXLS under Assumptions #1 through #5 of Key Concept 17. 17. 9Q > 0.x)ut. Suppose that Assumption #4 is replaced by the assumption that var(«.p.] 17.f.thcnE(fa) = 8x.43) with W ^ 0 6.) . Show that E(W2) < oo. conditional on Xx. that is... the law of large numbers.1 to prove that var(v() < co. d. implies that fa is consistent.. Derive the exact sampling distribution of the OLS estimator. u.. Show that E(ujiij\Xh Xj) = E(UJ\XJ)E(UJ\XJ) for i * j.. Show that ifE[(0-0)2] > Oas/i — * oo (that is.. Xj.. SXY P * &XY> where s^y is defined in Equation (3.d.. . and 9X > 0.Xn)^fa. What is the BLUE estimator of Bx? c.. X. LS e.Y/) are i.. b. Derive an expression for var(j§fi|Ai A. and Slulsky's theorem to show that the final term in the equation converges in probability to zero. Show that Vh(fa -fa) * N(0.

and fx is the marginal density of X ] b. however. 17.13 Consider the heterogeneous regression model Y/= BQi + BxiXj + «. where gXtY is the joint density of A and Y. between those two values. and Student t distributions.1.d.3)] are replaced by integrals.. then its probability is summarized by its probability density function (p.d.0 and (faj.4. cr2u).12 a. o-y^v).f. The probability density function of a normally distributed random variable (the normal p. where 60 and Bx are known positive constants. where E(fa) is the average value of BXl in the population. Show that E(eu\X = x) = e'{a+bx2) 17.1 T h e N o r m a l a n d Related Distributions M o m e n t s of C o n t i n u o u s R a n d o m and The Normal Distribution The normal distribution for a single variable. except that summations [for example.) is Variables This appendix defines and discusses the normal and related distributions. The variance is the expected value of (Y ~ fiy)2.The probability that Yfalls between two values is the area under its p. Suppose that H ~ N(0. Because probabilities cannot be negative. are restated here for convenient reference.2). Expected values and moments of continuous random variables. if Y is a continuous random variable.yx.d. the mathematical expressions for its probabilities involve integrals rather than the summations that are appropriate for discrete random variables. fy(y) . like those of discrete random variables. a + bx2). Because Y is continuous. Use the result in part (b) to show that E( Y\X = x) = a + • bx for suitably chosen constants a and b. fai) are distributed independently of A. given in Section 2.O\Y) a n d / i y ^ . (17. Show that E(eu) . Let j S * " denote the weighted least squares estimator. 4<\$em&t0& 17.). As discussed in Section 2.34) (17. Show that the density of Y given X = x can be written as \ ( y .35) ffy(y)dy.HY? = j'(y E(Yf)-A y J 2 / ^ ) ^ . Accordingly. the summation in Equation (2. where faj and Bu are random variables that differ from one observation to the next. a.692 CHAPTER 1 7 The Theory of Linear Regression with One Regressor a. The definitions of the chi-squared.i ( V j t Y l (17.33) where the range of integration is the set of values for which fY is nonzero. F.10). and the rth moment of a random variable is the expected value of r .(O-XYI°X)(X ~ Px)> [Hint: Use the definition of the conditional probability density fy\x=Ay) -= [gxAx< y)VUx(x)).-. Q = 60 + 0XX2. Use the result in part (a) to show that Y\X = x~ N(p. Suppose that E(UJ\XJ) . Let fY denote the probability density function of Y. where a and b are positive constants.f.f. We begin by presenting definitions of probabilities and moments involving continuous random variables. The probability that Y falls between a and b (where a < b) is PrCa < Y < b) = / fY(y)dy (17.p-y. c. fY(y) ^ 0 for all y.P*Y\X / The Normal and Related Distributions and Moments of Continuous Random Variables Probabilities and M o m e n t s of Continuous Random Variables 693 "*-*w = ^ W S M p ~2\ o-Y\x where a-y\x~ Vo-y(l .36) . Let >3pLS denote the OLS estimator of Bx given in Equation (17. the expected value of Y is E(Y)=txY= jyfy(y)dy. Suppose that var(^|. Pr( -oo s Y < oc) = 1.e K b. are probability-weighted averages of their values. which implies thal/^f y (y)rfy = l. Suppose that the conditional distribution of u given A = x is N(0.(T\/2l exp .32) Because Vmust take on some value on the real line. Thus var(Y) = E(Y. and (17. [Hint: See Equation (13.] b. Show that J3?LS -*-»• E(BX). Does fifLS -£-» E(fa)l Explain.

i.1 are true. General linear restrictions on B can be written as the q equations RB = r. but that Assumption #6 is not.41) holds and you do not know il ? What if you know fl? Construct an example of a regression model that satisfies the assumption £(uj\Xj) = 0. have exact tn_k_x and Fqjl-t:.1). then the GLS estimator is asymptotically more efficient than OLS. Does the result in Equation (18. by solving minfe [(Y .specifically. If in addition the errors are homoskedastic. 7 The TSLS estimator is a member of the class of GMM estimators of the linear model. where Xu is a binary variable that equals 1 if the Ith person is a female and X2i is a binary variable that equals 1 if the ith person is a male.5).til 730 CHAPTER 18 The Theory of Multiple Regression Summary 1. ut be uncorrelated with all observations on the regressors. When the errors are homoskedastic. Assumptions #1 through #5 are true. Assumptions #1 through #4 in Key Concept 18. and B. However. and if Cl is known or can be estimated. You are analyzing a linear regression model with 500 observations and one regressor. Assumptions #1 through #6 are true. If the error covariance matrix fl is not proportional to the identity matrix.5 Suppose that Assumptions #1 through #5 in Key Concept 18. not just with Xj as is required by OLS. but you think Assumption #6 might not be true (give two ways to construct the confidence interval).Xb)'Z]A[Z'(Y Xb)]. Write the model in the matrix form of Equation (18. B is the k + 1 vector of unknown parameters. where Y is the / i X l vector of observations on the dependent variable.1 are true. Define Y. an assumption that must be evaluated carefully in applications.1) in the matrix form of Equation (18.d.2 Exercises 18. in general. ditionally uncorrelated across observations and if E(uj\X) — 0. Explain how you would respecifiy the model to eliminate the perfect multicollinearity. X is the n X (k -+-1) matrix of n observations on the k -f-1 regressors (including a constant). b. Show that the columns of X are linearly dependent so that X does not have full rank. but you think Assumption #5 or #6 might not be true. The OLS estimator is B ^ (X'X)~lX'Y. j8 is consistent and asymptotically normally distributed. if the errors are homoskedastic and con5. The asymptotically efficient GMM estimator sets A = [£(ZjZ/«-)[ _ 1 . and normally distributed.2) for a hypothetical set of n — 5 observations. x distributions. then the conditional variance of J3 is var(faX) = crJ(A'A) -1 . 4.1 A researcher studying the relationship between earnings and gender for a group of workers specifies the regression model.31) hold? Explain. the asymptotically efficient GMM estimator in the linear IV regression model is TSLS. where A is a weight matrix. Explain how you would construct a confidence interval for fa if: a. 2. 18. In GMM. . respectively. Can you compute the BLUE estimator of B if Equation (18. The linear multiple regression model in matrix form is Y = XB f V. Write the regression in Equation (8. Yt = fa + Xufa • + • X2ifa + Uj. GLS requires that. B has an exact normal distribution and the homoskedasticity-only tand F-statistics. and U is the n X 1 vector of error terms.4 18. conditional on X.1 Consider the population regression of test scores against income and the square of income in Equation (8.1. the OLS estimator is efficient among linear conditionally unbiased estimators (OLS is BLUE). and this formulation can be used to test joint hypotheses involving multiple coefficients or to construct confidence sets for elements of B. but for which E(U\X) * 0„. A. Key T e r m s Gauss-Markov conditions for multiple regression (712) Gauss-Markov theorem for multiple regression (713) generalized least squares (GLS) (715) infeasible GLS (718) feasible GLS (718) generalized method of moments (GMM) (726) efficient GMM (727) heteroskedasticity-robust/-statistic (728) Exercises GMM /-statistic (728) mean vector (741) covariance matrix (741) 731 i Review the Concepts 18. 6. c. Under the first four least squares assumptions in Key Concept 18.3 18. 18. U. The Gauss-Markov theorem says that. the coefficients are estimated by making the sample covariance between the regression error and the exogenous variables as small as possible . 3. a. When the regression errors are i.

where 1>w is finite and positive definite. c Let Bx be the OLS estimator from the regression of Y on X (the restricted regression that excludes W). where A and W are. and assume that the assumptions in Key Concept 4.14) by OLS.28). Calculate the OLS estimates of fa. Then. . respectively. Show that the (1.6 Consider the regression model in matrix form. a..Wt'8. Y. c.and u(. Show that 0 < var«2) < oo. Now let fav be the "binary variable" fixed effects estimator computed by estimating Equation (10. r. a.732 CHAPTER 1 8 The Theory of Multiple Regression b.11) to derive the expressions for fa and 8X given in Key Concept 4. (ii) (Xj.d.fa + B:Xj + .2) and (18. b. Include all of the fixed effects in W Write out the matrix M^A. denote the z'th rows of X and W[as in Equation (18.. Let c be a nonrandom m x 1 vector.3)]. Consider the multiple regression model in matrix form Y = A/J — Wy + u. un... Derive an expression for E(UU') ~ il. Uj) have four finite. where A is an n X k{ matrix of regressors and Wis an nX k2 matrix of ml J •I . the OLS estimator B can be expressed p-=(X'MwX)-\X'MwY). b..are correlated. Suppose that c ^ 0m... = 0 at the 5% significance level. Suppose that Uj are i.5 Let Px and Mx be as defined in Equations (18.4 Consider the regression model from Chapter 4. Show that a.2. Y ~ A/} + Wy + U.2. in which the entity-specific sample means have been subtracted from X and Y.faWj + uit where for simplicity the intercept is omitted and all variables are assumed to have a mean of zero. --. 18. nonzero moments. regressors. 81.4. as shown in Exercise 18.1.. a. Assume that (i) E(ut\Xh Wj) .for all i and. d.13) is equal to the expression for cri given in Key Concept 4.5u M -+• Uj for i = 2. Derive Equations (18. Calculate the R2 of the regression. Use the general formula for /\$ in Equation (18. b.are correlated. b.i. Show that var(Q) . Dlit D2. 18.. (iii) (A. a.BXM. Use the expression for B given above to prove that fav —. with the till a. Show that Assumptions #1 through #4 in Key Concept 18.) are i. If Wj and «..d.3 Let W be an m X 1 vector with covariance matrix 2jy. What are R..24) and (18.1 hold.) 18. Y.2U).Test the hypothesis that 0. . Provide conditions under which Bx has a smaller asymptotic variance than 8\.. Explain how to test the null hypothesis that the relationship between test scores and income is linear against the alternative that it is quadratic. Explain how to estimate the model by GLS without explicitly invert ing the matrix ft.c' 2>yC. Whether or not Wj and «.are correlated.3 hold. and (iv) there is no perfect multicollinearity. Write the model in the matrix form given in Equations (18.u.11) by OLS and let faM be the "de-meaning" fixed effects estimator computed by estimating Equation (10.9 This exercise shows that the OLS estimator of a subset of the regression coefficients is consistent under the conditional mean independence assumption stated in Appendix 7. (Hint: Transform the mode] so that the regression errors are ux.7 Consider the regression model.i. Calculate s2u.1) element of 2 ^ in Equation (18. with mean 0 and variance 1 and are distributed independently of A. 18.] 18.4).27) and (18.11) using a full set of fixed effects. and q'l 18. and let Q = c'W. . n. allowing for the possibility that W.1 are satisfied.. where ux = ux and u-t — 0. 18. where S is a k2 X 1 vector of unknown parameters. nXkx and nxk2 matrices of regressors. Wj. Let A/ and W.25).17. b... These are Assumptions #l-#4 of Key Concept 18. Prove that PxMx ~ 9nKn and that Px and Mx are idempotent. then fa is inconsistent. Suppose that all six assumptions in Key Concept 18. and fa. 18. Write the null hypothesis in the form of Equation (18. Y.. Wt.) but Wt and u{ might be correlated and let fa and fa be the OLS estimators for this model. = BxXj I.8 Consider the regression model Yj = fa + BxXj + ut. Suppose that Xj is distributed independently of ( W h «.2 Suppose that a sample of n = 20 households has the sample means and sample covariances below for a dependent variable and two regressors: Exercises 733 A .. fa >i ' fa PIb.3. u2. DUJ and no constant term. [Hint: Write Equation (10.

) b. Show that ^lUvt 18.36) is equivalent to the homoskeskasticityonly F-statistic in Equation (7.) Consider the panel data model Y-j = (3Xj. Wx is the « X l vector with Ith element WXi.yq •= yCq . c Show that E Exercises b. #2.1.Yj = MY„ Xj = MXj. Show that (Y . i. Let W = [1 Wx W2 ••• Wr]. a. (Hint: Mis idempotent. a. where all variables are scalars. and #4 in Key Concept 10. Consider the regression described in Equation (12. where i is a T X 1 vector ones.AA'. + atj + ui(. (Hint: Note that Cq = yq implies 0 = Cq . Show that Fin Equation (18.WS. Use your answers to (b) through (d) to prove Equation (10.(Y-Xfa(Y ~ r).14 Consider the regression model Y .j. Show that d' Cd > 0.12 a.• = XiT)\Uj = (uil uiT) '. Show that C .3 hold and strengthen Assumption #3 so that Xj. and uit have eight nonzero finite moments. where R is q X (k +1) with rank q. Show that trace(C) = rank(C).show that VnT(p . c Let d be a n X 1 vector.] c. Show that assumptions (i) and (ii) imply that E(U\X. Let o^c/H^rerf be the infeasible clustered variance estimator computed using the true errors instead of the residuals so that cr^^. Show that fa(X'X)fi (Rfa'[R(X'Xy1R]~1(Rfa. Use the expression for B given in Exercise 18.o^Q2x). that is.-) ^UX-Uj. y) (Y~Xb)'(YXb) + y'(Rb r). Show that the method for computing the /-statistic described in Key Concept 12. and so forth.6 (using a homoskedasiticity-only F-statistic) and the formula in Equation (18. and Exercise 18.^ 2xx . where 1 is an n X 1 vector of ones. Let B be the value of b that solves the constrained minimization problem. 18.v.X f a ' ( Y .B = (n-1X'MwX)-\n-)X'MwU). where 2 ^ = E(XjXj').fa --^ N(0. that pEff.^ A if An jj —*-» A y for all i.cius. ii. where AnJj and Atj are the (ij) elements of A„andA.XB + XJ.yq = CCq . Assume that Assumptions #1. Ir).17). h -*-» B.10 Let C be a symmetric idempotent matrix.MUj. S h o w t h a t ^ B = (2-=1 A/A.734 CHAPTER 1 8 The Theory of Multiple Regression conditional mean independence assumption (i) replacing the usual condi tional mean zero assumption. Let (jTSLS denote the vector of two stage least squares residuals. under conditions (i) through (iv). (b. Let M=I7 T'1 i . Show that WUTSLS = 0. 2 V'CV-xl GMM 18.13 Consider the problem of minimizing the sum of squared residuals subject to the constraint that Rb = r. Uj -. [Hint: Use the results in (a). where Xx has kx columns and A2 has k2 columns.] 18.3 can be written as J3 = ( S ^ A / A ^ S t i A / % b.y2q .65).66) is the solution to Equation (18. Show that P = f\$. W) .~ >«'<*. and so forth. Also let Ff = (YiX Y.25). For the asymptotic calculations in this problem.63) produce the same value for the /-statistic.) Qx. f.13. Let 7}j .(X'X)-1R'[R(X'X)-1R']-1(Rp c.r)'[R(X'X)~ R']-i(RP l 735 . a.^ b. Show that the Lagrangian for the minimization problem is L(b. b.yq =. Partition A as [A t A 2 ] and B as [fa' B{].15 (Consistency of clustered standard errors.6 to write B . e.that is. Show that the fixed effects estimator of 8 from Section 10. where A is n X r with A'A = / r (//«£• C is positive semidefinite and can be written as QAQ' as explained in Appendix 18. a.2 A W 2 ^ V 2 W T .yq and solve for y. Show that the eigenvalues of C are either 0 or 1. A/J) = .13). 0 t l X t J.xle. "» Xm -*• c l^tQ^r-^A/A^andQi^^ d. Show that n^X'M^X . 18.ered o-l __^i^ . Show that J GMM d pW-GMM) -*-> 0. L e t f i = [/A. — ^ M0. 2XW = E(XjWj'). <r2). nxX'MyjU . Show that A' V ~ N(dr. i). and {Xn *r.Xj'Uj/Vf and a = varfa).^ = 2 2 nf^U(XiUj) . 18. Use (c) and the law of iterated expectations to show that 0. d.GMM i n Eq U a t ion (18.) b. [The matrix An .X f a (RP . a. d. Suppose thatA 2 'Y = 0 ^ x l . Show that p ffis the efficient GMM estimator . '*.r). Show that o. Use (a) through (d) to conclude that. where y is a q X 1 vector of Lagrange multipliers. suppose that T is fixed and n > oo a.11 Suppose that C is an n X n symmetric idempotent matrix with rank r and letK-MO^). b. Show that Vn(frE*GMMc.

is the ith element of 6. I.= l|A. Suppose that I. Show that the OLS estimator can be written as P-[%liXiXj) {^IjXjY^B^i^IjXjX. that is. Is 0 unbiased? Is 0 consistent? Explain. i. The purpose of this appendix is to review some concepts and definitions from a course in linear algebra.Jjs&- . clus. Let It denote a binary random variable that indicates the nonmissing observations. Show that XX WX (X'MwX)~l iWWy WX(X'MwXyl l -1 A'A WX X'W WW -i X'Y WY . . .) = p. matrices.«. ut\ are i. collected either in a column (a column vector) or in a row (a row vector).] 18. = 1 when Yt > 0. Is /3 unbiased? Is £ consistent? Explain. that is. c Show that £ = (X'Mwxyl XMWY. Suppose that the probability that the ith observation is missing depends on both Xj and ut. Definitions of Vectors and Matrices A vector is a collection of n numbers or elements. . that is. iii. Pr(^ = l|A. and the elements of matrix algebra used in Chapter 18. Let X ~ MWX and Y = MWY. = 1 if observation i is not missing and /.ereti -£— 0. 18.Show that B is unbiased and consistent.17 Consider the regression model in matrix form Y = XB f Wy f U. Show that j8 is unbiased and consistent. where all variables are scalars and the constant term/intercept is omitted for convenience.= 0 if observation i is missing. . = 1JA. ut) = p(X{.16 This exercise takes up the problem of-missing data discussed in Section 9. Xj. Suppose that 8 = 1 and that A. where Mw^i~w(wwylw.^ (j^hXjUj WvPA 18. Suppose that the probability that the ith observation is missing depends of A . ut). not to replace such a course.t the least assumptions in Key Concept 4.°^... Suppose tha.. Show that the OLS estimators of B and y can be written as = Y_ b. b.. d. and ut are mutually independent standard normal random variables [so that both Xt and ut are distributed N(0.2) says that /3 ^ (A'A)" 1 X'Y. a boldface symbol denotes a vector or matrix.d. u . where X and W are matrices of regressors and R and y are vectors of unknown regression coefficients. Assume that {/. The Frisch-Waugh theorem (Appendix 6. n.-. but It . ii. Show that a\ ctustered — ^ cr2 [Hint: Use an argument like that used to show Equation (17. Suppose that data are "missing completely at random" in the sense that Pr(/.. Now suppose that some of the observations are missing. c. Let « t = Yj ~ pXi and a\ clus[ered = „ ^ ^ H ( ^ ! « « ) 2 [this is Equation (10. where p is a constant. .0 when Yj < 0. ) =/?(A}). Consider the regression model Yj = X{B + u„ i = 1 .The /i-dimensional column vector b and the rc-dimensional row vector c are vm and c = [c| c2 6= b2 where b: is the first element of b and in general 6.1)]. Throughout.3 are satisfied. Use the result in (c) to prove the Frisch-Waugh theorem.Summary of Matrix Algebra 736 CHAPTER 1 8 The Theory of Multiple Regression g..16) to show that o\ chislered 737 a. Algebra This appendix summarizes vectors.but not on u{.27) in matrix form]. Show that the least squares estimator of B is unbiased and consistent. Pr(/ .2. iv.i. a. then use your answer to (f).1 S u m m a r y of M a t r i x X'W -1 WW -(X'MWX)-1X'W(W'W)^ (W'Wy1W'X(X'MwXy1X'W(W'Wy1 (WW) + (Hint: Show that the product of the two matrices is equal to the identity matrix). .