Questions and Answers on Methods for Time Series - Cross-Section Data

L. Magee 1. To estimate in the model yit = i + xit +

Fall, 2008

, where i = 1, . . . , n , and t = 1, .., T

where xit does not include a constant term, three common estimators are b: pooled OLS, with i = for all i RE : random eects estimation, with i treated as a random variable F E : xed eects estimation, with i treated as regression coecients For each estimator, describe assumptions under which it would be preferred to the other two, and how to test these assumptions. 2. Consider the model yit = i + xit + i 1 1 2 2 3 3 t 1 2 1 2 1 2 xit 0 1 0 1 0 1 yit 11 12 8 9 18 20
it ,

where i = 1, 2, 3 and t = 1, 2. The data are

(a) Compute the xed eects estimate of (b) With these data, will the value of the xed eects estimate be very dierent from the OLS estimate? (Computing OLS is not required.) Briey explain your answer. 1 0 0.5 0 0 4 0 2 3. (a) Find two symmetric 22 matrices A and B such that the matrix 0.5 0 0.5 0 0 2 0 2 can be written as A B, the Kronecker product of A and B

 (b) Use the property (A B)1 = A1 B 1

0 0.5

0 4 to get the inverse of 0.5 0 0 2

0 2 0.5 0 0 2

(The inverse of a 2 2 matrix is: a11 a12 a21 a22

= (a11 a22 a12 a21 )1

a22 a21

a12 a11

4. Consider the time-series/cross-section regression model yit = i + xit +

, where i = 1, . . . , n , and t = 1, . . . , T

where xit does not include a constant term. Let dier across i. Then the constant term i can be specied as part of the i vector. The model becomes yit = xit i +
it

, where i = 1, . . . , n , and t = 1, . . . , T

When n is not large, the i vectors are often estimated by the seemingly unrelated regression equations (SURE) method. What assumptions are made about the model? 5. Write the variance-covariance matrix of the disturbance terms ((i ) + eects model yit = + xit + ((i ) +
it ) it ) it s

in the standard SURE

in the random

making the usual random eects model assumptions. 6. (18 marks: 4 for a, 4 for b, 6 for c, 4 for d) Consider the estimation of in the model yit = i + xit +
it

, where i = 1, . . . , n , and t = 1, .., T

where xit does not include a constant term. xit and are k 1 vectors. (a) (i) Describe the pooled OLS estimator of . (ii) Under what conditions might the pooled OLS estimator be a good choice? (b) (i) In the random eects model, what assumptions are made about the i terms? (ii) Describe the variance-covariance matrix of the disturbance terms in the random eects model resulting from these assumptions. 2

(c) (i) Express the xed-eects estimator of as a function of the data, yit and xit , and the group means, yi and xi . (ii) Describe the characteristics of a regressor variable that result in an inability to estimate its coecient by the xed-eects method, even though it could be estimated by the other methods. (d) When n is small, it is possible to estimate the seemingly unrelated regression equations (SURE) model yit = i + xit i +
it

, where i = 1, . . . , n , and t = 1, .., T

(i) Describe the assumptions that usually are made about the means, variances, and covariances of the disturbance terms
it

in the SURE model.

(ii) Express, using Kronecker product notation, the variance-covariance matrix of the disturbances that follows from these assumptions. 7. A and B are 1 2 matrices, with A = [ 1 2 ] and B = [ 1 1]. Write the numerical values of the following matrices. (a) A B B B)B

(b) (AA ) (c) (A

8. The the n-equation seemingly unrelated regression equations (SURE) model, where the regressors are the same in each equation, is y1 = X1 + y2 = X2 + . . . . . . yn = Xn + where yi and
i 1 2

are T 1 vectors, X is a T k matrix, and the i s are k 1 coecient vectors.

(a) In stacked form, this model is written as y = X + . State the dimensions of y, X , , and , and write them as functions of the variables appearing in the model as originally written. (b) Assuming E
i

= 0 and EX

= 0 for each i, and E

i j

= ij I for all i, j, write the (that is, = E )

variance-covariance matrix of the stacked-model disturbance vector using Kronecker product notation. 3

(c) Write X as a function of X using Kronecker product notation. (d) The SURE estimator of is the feasible GLS estimator = X 1 X X 1 y. b1 b = 2 , where bi = (X X)1 X yi . Prove, for this special case of the SURE model, that . . . bn 9. Consider the xed-eects model yit = i + xit +
it , 1

where i = 1, 2, 3 and t = 1, 2.

(a) Write a formula for the xed-eects estimator of . (b) Compute the numerical value of the xed eects estimate of for the data i t xit yit 1 1 2 2 3 3 Answers 1. Use b if it appears that i = j for all i, j, or if they are close to being equal relative to the size of the standard errors of their estimates. For example, use b if an F -test of H0 : i = j for all i, j is accepted. Use RE if there is evidence that (i) the i s dier from zero and (ii) the i s are not
2 2 correlated with xit . Check (i) by testing = 0 in the random eects model, where is the

1 2 1 2 1 2

0 1 0 1 0 -1

11 12 8 9 28 24

variance of the random eect i , and check (ii) by a Hausman test, where H0 is that the correlations between i s and the elements of xit are zero. (The Hausman test interprets this RE ) = plim(F E )). as H0 : plim( Use F E if the i s are not equal (e.g. if H0 : i = j for all i, j is rejected) and if the i s are correlated with xit (e.g. if the Hausman test rejects). 2. (a) x1 = 0.5, x2 = 0.5, x3 = 0.5, y1 = 11.5, y2 = 8.5, y3 = 19. The group-mean-dierenced data are then

i 1 1 2 2 3 3

t 1 2 1 2 1 2

xit xi -0.5 0.5 -0.5 0.5 -0.5 0.5

yit yi -0.5 0.5 -0.5 0.5 -1 1

The xed eects estimator is F E =

xi )(yit yi ) .25 + .25 + .25 + .25 + .5 + .5 = = 1.333 (xit xi )2 6 .25 it

(b) F E equals OLS. Because the xi s are all the same, they all are equal to the full-data x. Therefore the mean-dierencing that occurs in the OLS calculation, which subtracts x is the same as the within-group dierencing that is done in xed-eects estimation. (Aside: The yi s are not equal, but that doesnt aect F E since it (xit xi )(yit yi ) = it (xit xi )yit the rst-dierencing of yit in the formula is just for interpretations sake.) Another way to look at it is to consider the xed eects dummy variables in the D matrix that appears when the model is written in LSDV form. The equality of the xi s shows that the columns of D are orthogonal to the xit variable. Changing from OLS to FE amounts to including D in the set of regressors, and this has no eect on the estimate of . This orthogonal-regressor property is a special case of the Frisch-Waugh theorem. 3. (a) A = 2 1 1 1 and B = 0.5 0 0 2 , or A and B where

A = kA and B = k 1 B, for any k = 0. (b) A1 = 1 1 1 2 and B 1 = 2 0 0 .5 Then (A B)1 = A1 B 1 = 1 1 1 2 2 0

0 2

0 .5

0 .5 0 .5 = 2 0 4 0 0 .5 0 1

4.

mean: E

it

=0
it jt

allows correlation across cross-section units at the same time period: E 5

= ij

 equal variance within cross-section units but not equal across cross-section units: E no autocorrelation: E
it is

2 it

= ii

= 0 for all t = s
it js

no correlation across both time and cross-section unit: E 5. The assumptions are E
it

= 0 for all i = j and t = s

=0;

E(i ) = 0 ;

2 it

= 2 ;

2 E(i )2 = ;

it (i

) = 0

Express the ordering of the observations in the usual way, with the time series t = 1, . . . , T for cross-section unit i rst, followed by T observations for i = 2, etc. Then the variance-covariance matrix is A 0 0 ... 0 0 . . . . . . =I A

0 A 0 ... . .. . 0 . . . . .. . . . . . 0

0 ... ... A

where A=
2 ( 2 + ) 2 . . . . . . 2 2 2 . . . 2 2 . . . . . .

2 = 2 I +

2 2 ( 2 + ) . . . .. 2 . . .. . . . 2

2 . . . . . . ( 2 + )

xi1

yi1

, i = 1, . . . , n.

xi2 . . . 6. Dene the T k matrices Xi = x it . . . xiT

yi2 . . and the T 1 vectors yi = . y it . . . yiT

 Then dene the nT k matrix X =

X1 X2 . . . Xi . . . Xn

y1

y2 . . . and the nT 1 vector y = y . i . . . yn

Let represent a column vector of ones, having whatever number of elements is necessary to conform with the other matrices and vectors in the expression. (a) (i) The pooled OLS estimator of is the slope coecient part of the OLS estimator in the P model yit = + xit + it , where the i s are all set equal to . That is, = P (X X )1 X y, where X = [ X]. Another way to write P is P = ( x= (nT )1
it xit it (xit

x)(xit x) )1 (nT )1
it yit .

it (xit

x)(yit y ), where

(a k 1 vector), and y =

(ii) The pooled OLS estimator is a good choice when the restriction i = j for all i, j is plausible. (b) (i) The assumptions (including the ones about E
it it )

are
it (i

= 0 ; E(i ) = 0 ; E

2 it

2 = 2 ; E(i )2 = ; E

) = 0

(also assume no autocorrelation across t in the s, s, or their cross-products. (ii) A 0 0 ... 0 0 ... 0 0 A . . . . . =I A . 0 .. . . .. . . . . . . . . . 0 where A=
2 ( 2 + ) 2 . . . . . . 2 2 2 . . . 2 2 . . . . . .

0 ... ... A

2 = 2 I +

2 2 ( 2 + ) . . . .. 2 . . .. . . . 2

2 . . . . . . ( 2 + )

(c) (i) F E = (

it (xit xi )(xit xi )

)1

it (xit xi )(yit yi ), 7

where xi = T 1

t xit

(a k 1

vector), and yi = T 1

t yit .

(ii) If one of the elements of xit does not vary with t for any xed i, then that element of xit xi will equal zero for every i and t (e.g. if the variable is years of education, and this variable does not change over time for any individual i in the sample). The matrix being inverted in the expression for F E will not have full rank and is not invertible. But as long as there is variation across i (e.g. the individuals do not all have the same number of years of education), then there is variation in xit x so that this problem is P . The random eects estimator does not use group-mean dierencing, not present for so it would still work too. (d) (i) mean: E
it

=0
it jt

allows correlation across cross-section units at the same time period: E E

= ij

equal variance within cross-section units but not equal across cross-section units: = ii
it is

no autocorrelation: E t=s 11 12

= 0 for all t = s
it js

no correlation across both time and cross-section unit: E

= 0 for all i = j and

. . . 1i

. . . 1n

21 22 . . . 2i . . . 2n . . . . . .. . . . . . . . . . . . (ii) Let = i1 i2 . . . ii . . . in . . . . . .. . . . . . . . . . . . n1 n2 . . . ni . . . nn

Then the disturbance variance-covariance matrix, when the observations are ordered in the same way as described earlier, is 7. (a) A B =[ 1 1 2 2 ] B=5 B)B = [ 1 1] = [ 5 5] 1 2 [ 1 1] 1 1 = 1 1 2 2 1 1 = 2 4 I, where I is a T T identity matrix.

(b) (AA ) (c) (A 8. (a)

 y = X

y1 y2 . . . yn X 0 . . . 0 1 2 . . . n

, an nT 1 vector 0 X 0 0 ... 0 0 . . . , an nT nk matrix

0 ... .. . 0

... X

, an nk 1 vector

2 . , an nT 1 vector = . .
n

11

12

13

. . . 1n

21 22 23 . . . 2n (b) = I, where I is a T T identity matrix and = . . .. . . . . . n1 n2 n3 . . . nn (c) X = I X where I is an n n identity matrix (d) First consider each multiplicative factor of , one at a time. X 1 X = (I X )(1 I)(I X) = 1 X X so that (X 1 X )1 = (X X)1 Also, X 1 y = (1 X )y

Multiplying these two terms and simplifying gives = (X 1 X )1 X 1 y = ( (X X)1 )(1 X )y = (I (X X)1 X )y (X X)1 X 0 0 (X X)1 X = . . . 0 (X X)1 X y1 (X X)1 X y2 = . . . 1 X y (X X) n 9. (a) F E =
it (xit

...

0 0 . . .

y1 y2 . . . yn

0 ... .. . 0

. . . (X X)1 X

xi )(yit yi ) = 2 it (xit xi ) and yi = T 1

it (xit

xi )yit 2 it (xit xi )

where xi = T 1

t xit

t yit

(b) x1 = 0.5, x2 = 0.5, x3 = 0.5, y1 = 11.5, y2 = 8.5, and y3 = 26. Then i 1 1 2 2 3 3 t 1 2 1 2 1 2 xit xi -.5 .5 -.5 .5 .5 -.5 yit yi -.5 .5 -.5 .5 2 -2 SUMS so that F E = 3/1.5 = 2.0 (xit xi )(yit yi ) .25 .25 .25 .25 1 1 3 (xit xi )2 .25 .25 .25 .25 .25 .25 1.5

10