Geneva School of Economics and Management University of Geneva

Institute of Economics and Econometrics Spring 2017

Advanced Econometrics II Professor Jaya Krishnakumar

Problem Set 2 (Adv. Econometrics I - Exam Jan17)

Exercise 1 (Theory)
We postulate the following model with a change of regime from period one to two :

yi1 = α + εi1 i = 1, .., N (1)

yi2 = β + εi2 i = 1, .., N (2)

The stochastic assumptions are

E(εi1 ) = E(εi2 ) = 0
E(ε2i1 ) = E(ε2i2 ) = σ 2
E(εik , εjs ) = 0 for i = 1, . . . , N ; j = 1, . . . , N ; k = 1, 2; s = 1, 2

1. Write the model in the following matrix form

(
y1 = X1 α + ε1
y2 = X2 β + ε2

Specify the content and dimension of all the matrices.

2. Regroup the model in point 1 as follows

Y = Xθ + ε (3)

Specify the content and dimension of all the matrices and show that X can be written as a
Kronecker product.
3. Calculate E(ε) and V (ε).
4. Give the Ordinary Least Squares (OLS) estimator of θ and deduce the OLS estimators of
α and β .
5. Calculate V (θ̂OLS ).

Suppose now that we have correlated errors for the same individual across periods but not
between dierent individuals. The new stochastic assumptions are

E(εi1 ) = E(εi2 ) = 0 ∀i
 2

 σ i = j, t=s

0 i 6= j, t=s
E(εit , εjs ) =


 σ12 i = j, t 6= s
i 6= j, t 6= s

0
Advanced Econometrics II, February 2017 2

6. Give the variance-covariance matrix of the model under these new assumptions. Show that
it can be written as a Kronecker product.
7. What is the best estimator of θ denoted as θ̃ ?
Give its expression and show that it is equivalent to what you found in point 4.
8. Calculate V (θ̃). Comment.

Exercise 2 (Theory and empirics)

Jerey Kling studies the impact prison sentences' length on employment and earnings following
prison release in his paper "Incarceration Length, Employment, and Earnings" (J. R. Kling.
2006. American Economic Review, 96, 863-876).
Table 1 below provides a possible set of results from such a study. The author proposes three
dierent dependent variables to measure the earnings of the defendant upon release :
1. y is dened as the fraction of the calendar quarters with any positive earnings.
2. y is dened as the fraction of the calendar quarters with earnings above the poverty line.
3. y is dened as the average quarterly earnings.
Columns 1,2 and 3 of Table 1 summarize the results for these three dependent variables res-
pectively. Kling also considers two estimation strategies. In the rst row, we report the coe-
cient estimate of prison sentence from regressing the dependent variable by the Ordinary Least
Squares (OLS) on prison sentence and other controls. In the second row, we show the coecient
estimate of prison sentence from regressing the dependent variable by 2-Stage Least Squares
(2SLS) on prison sentence and other controls. Finally, heteroskedasticity-robust standard errors
are displayed in parenthesis below their associated coecient estimates.
Table 1  OLS and Instrumental Variables estimates of eects of an additional year of incar-
ceration on labor market outcomes nine years after incarceration began
Earnings > zero Earnings > poverty Avg. earnings
OLS : coecient of -0.0031 -0.0038 -44
prison sentence (0.0010) (0.0018) (16)
2SLS : coecient of 0.0169 0.0069 248
prison sentence (0.0067) (0.0022) (114)

1. Give a brief economic and statistical interpretation of the estimation results reported in
Table 1.
2. Explain why endogeneity might be an issue in their regression. Give two specic examples
of why the length of a former inmate's prison sentence might not satisfy the exogeneity
assumption.
3. For each example above, indicate whether you expect the OLS estimates to be biased towards
or away from zero.
To solve the potential endogeneity issue, Kling proposes the following instrumental variable
strategy. In the United States, judges are assigned randomly to federal cases. Moreover, some
judges are inherently more indulgent than others in their sentencing decisions even when faced
with identical cases. The author therefore proposes to use a measure of judges' leniency as an
instrument for the length of the defendant's prison sentence.
Advanced Econometrics II, February 2017 3

4. Do you think that judges' leniency is a good instrument for the duration of the defendant's
prison sentence ?
Give the conditions needed to ensure that the instrumental strategy provides consistent
estimates and comment on whether these conditions are likely to be satised in this example.
5. Compare the results from the 2SLS regression with those from the OLS regression reported
in Table 1.

Measuring precisely the leniency of a judge is dicult and can lead to measurement error in
the variable that Kling wants to use as an instrument.

Suppose that the relationship between the prison inmate future earnings yi and the length
of the prison sentence xi is given by :

yi = β0 + β1 xi + εi i = 1, ..., n (4)

where the duration of prison sentence is endogenous such that E[εi |xi ] 6= 0. In his paper, Kling
proposes to use as a valid instrument judges' leniency, zi∗ .
Unfortunately, we suppose that we do not observe zi∗ directly but only zi :

zi = zi∗ + vi (5)

with the following assumptions on the measurement error E(vi |zi∗ ) = 0 and E(εi |zi∗ ) = 0.
6. Write equation (4) in matrix notation : y = Xβ + ε. Give the content and dimensions of all
the matrices.
The instrumental variable estimator can be written as
 IV 
β̂0
β̂IV = = (Z 0 X)−1 Z 0 y (6)
β̂1IV

where  
1 z1
Z = 1 z2 
 
.. ..
. .

7. Calculate the limit in probability of β̂ IV . Is β̂ IV a consistent estimator of β ?

Make all the necessary assumptions.
8. Give the conditions on the relationship between v and ε under which β̂ IV a consistent
estimator of β .
9. Suppose we implement a 2SLS procedure. The rst-stage regression of this procedure will
be given by :
xi = α0 + α1 zi∗ + ei (7)
(a) Given (5), what can we say about the OLS estimators of α0 and α1 ?
Justify your answer.
(b) Combining your answer to (a) with your answer to 8., give a conclusion on the consis-
tency of 2SLS when the instruments are measured with error.
Advanced Econometrics II, February 2017 4

Exercise 3 (Empirics)
In the beginning of the 80's, China was accused by the United States to practice dumping on its
exports of barium chloride to the U.S. (by selling at an unfairly low price). In October 1983, the
U.S. barium chloride industry led a complaint to the International Trade Commission (ITC)
which was ruled in its favor.
Krupp and Pollard (1996) propose to analyze the eects of anti-dumping lings by the U.S.
barium chloride producers on the imports of barium chloride from China. We reproduce some
of their results to investigate whether the Chinese imports dier noticeably before and after
the ling.
Using monthly data from February 1978 through December 1988, we estimate the following
regression :

log(chnimp)t = α0 + α1 log(chempi)t + α2 log(gas)t + α3 log(rtwex)t

(8)
+α4 bef ile6t + α5 af f ile6t + α6 af dec6t + ut

with

chnimp : the volume of U.S. imports of barium chloride from China.

chempi : an index of chemical production to control for the overall demand of barium.
gas : the volume of gasoline production.
rtwex : an exchange rate index reecting the strength of the dollar against other currencies.
bef ile6 : a dummy variable equal to one during the six months before the ling.
af f ile6 : a dummy variable equal to one during the six months after the ling.
af dec6 : a dummy variable equal to one during the six months after the positive decision.

To answer to the following questions, refer to the results in appendix.

We estimate equation (8) by Ordinary Least Squares (OLS).

1. Give a brief economic and statistical interpretation of the estimation results.
Is there evidence that Chinese imports were unusually high before the ling ?
2. Compute the percentage change in the Chinese imports after the decision in favour of the
United States.
3. Propose a test to examine whether the joint impact of all the explanatory variables, except
the constant term, is signicant. Specify the null and the alternatives hypotheses as well as
the test and its conclusion.
Since we are dealing with time series data, we are worried that seasonality might be an issue
in our OLS estimation. We decide to re-estimate equation (8) adding seasonal dummies.
4. Comment on the new estimation results. Are the higher Chinese imports just before the
suit was led due to seasonality ? Justify your answer.
5. Test the joint signicance of all the seasonal dummies. Specify the null and the alternatives
hypotheses as well as the test. What do you conclude ?
6. Test whether there is rst-order autocorrelation in the error term. The coecient of rst-
order autocorrelation is 0.2707.
Advanced Econometrics II, February 2017 5

7. Explain the dierent steps of the procedure that we will use to test third-order autocorre-
lation in the error term (see the output in appendix). What is the conclusion of the test ?
8. If the null hypothesis is rejected in the above question, suggest a transformation to correct
for autocorrelation of order 3, i.e. AR(3), and give the expression of the transformed y.
A. Appendix

OLS Regression
reg lchnimp lchempi lgas lrtwex befile6 affile6 afdec6

Source | SS df MS Number of obs = 131

-------------+----------------------------------

Model | 19.4051456 6 3.23419093

Residual | 44.2471061 124 .356831501 R-squared = 0.3049

-------------+---------------------------------- Adj R-squared = 0.2712

Total | 63.6522517 130 .489632706 Root MSE = .59735

------------------------------------------------------------------------------

lchnimp | Coef. Std. Err. t P>|t| [95% Conf. Interval]

-------------+----------------------------------------------------------------

lchempi | 3.1172 .479202 6.50 0.000 2.168725 4.065675

lgas | .1963049 .9066233 0.22 0.829 -1.598157 1.990766

lrtwex | .9830093 .4001536 2.46 0.015 .1909934 1.775025

befile6 | .0595742 .26097 0.23 0.820 -.4569584 .5761069

affile6 | -.0324067 .2642973 -0.12 0.903 -.5555252 .4907118

afdec6 | -.5652446 .2858353 -1.98 0.050 -1.130993 .0005036

_cons | -17.80195 21.04551 -0.85 0.399 -59.45692 23.85302

------------------------------------------------------------------------------
Advanced Econometrics II, February 2017 6

OLS Regression with seasonal dummies added

reg lchnimp lchempi lgas lrtwex befile6 affile6 afdec6 feb mar apr may jun jul
aug sep oct nov dec

Source | SS df MS Number of obs = 131

-------------+---------------------------------- F(17, 113) = 3.71

Model | 22.8083791 17 1.34166936 Prob > F = 0.0000

Residual | 40.8438726 113 .3614502 R-squared = 0.3583

-------------+---------------------------------- Adj R-squared = 0.2618

Total | 63.6522517 130 .489632706 Root MSE = .60121

------------------------------------------------------------------------------

lchnimp | Coef. Std. Err. t P>|t| [95% Conf. Interval]

-------------+----------------------------------------------------------------

lchempi | 3.265067 .4929297 6.62 0.000 2.288485 4.24165

lgas | -1.278206 1.389015 -0.92 0.359 -4.030095 1.473683

lrtwex | .6630341 .471303 1.41 0.162 -.2707021 1.59677

befile6 | .1397036 .2668075 0.52 0.602 -.3888901 .6682973

affile6 | .0126343 .2786866 0.05 0.964 -.5394941 .5647627

afdec6 | -.5213008 .3019498 -1.73 0.087 -1.119518 .0769161

feb | -.417716 .3044432 -1.37 0.173 -1.020873 .1854408

mar | .0590529 .2647304 0.22 0.824 -.4654258 .5835316

apr | -.4514835 .2683861 -1.68 0.095 -.9832049 .0802378

may | .0333114 .2692426 0.12 0.902 -.5001067 .5667294

jun | -.2063286 .2692517 -0.77 0.445 -.7397649 .3271077

jul | .0038404 .2787666 0.01 0.989 -.5484466 .5561274

aug | -.157059 .2779935 -0.56 0.573 -.7078142 .3936962

sep | -.1341598 .2676556 -0.50 0.617 -.6644339 .3961142

oct | .051691 .2668511 0.19 0.847 -.4769892 .5803712

nov | -.246259 .2628271 -0.94 0.351 -.7669669 .2744489

dec | .1328415 .2714237 0.49 0.625 -.4048978 .6705809

_cons | 16.78074 32.4288 0.52 0.606 -47.46656 81.02804

------------------------------------------------------------------------------
Advanced Econometrics II, February 2017 7

Procedure to test AR(3) in the error term

1)

reg lchnimp lchempi lgas lrtwex befile6 affile6 afdec6

Source | SS df MS Number of obs = 131

-------------+---------------------------------- F(6, 124) = 9.06

Model | 19.4051456 6 3.23419093 Prob > F = 0.0000

Residual | 44.2471061 124 .356831501 R-squared = 0.3049

-------------+---------------------------------- Adj R-squared = 0.2712

Total | 63.6522517 130 .489632706 Root MSE = .59735

------------------------------------------------------------------------------

lchnimp | Coef. Std. Err. t P>|t| [95% Conf. Interval]

-------------+----------------------------------------------------------------

lchempi | 3.1172 .479202 6.50 0.000 2.168725 4.065675

lgas | .1963049 .9066233 0.22 0.829 -1.598157 1.990766

lrtwex | .9830093 .4001536 2.46 0.015 .1909934 1.775025

befile6 | .0595742 .26097 0.23 0.820 -.4569584 .5761069

affile6 | -.0324067 .2642973 -0.12 0.903 -.5555252 .4907118

afdec6 | -.5652446 .2858353 -1.98 0.050 -1.130993 .0005036

_cons | -17.80195 21.04551 -0.85 0.399 -59.45692 23.85302

------------------------------------------------------------------------------

predict uh, residuals

Advanced Econometrics II, February 2017 8

2)

reg uh lchempi lgas lrtwex befile6 affile6 afdec6 L1.uh L2.uh L3.uh

Source | SS df MS Number of obs = 128

-------------+---------------------------------- F(9, 118) = 1.72

Model | 5.03366421 9 .559296023 Prob > F = 0.0920

Residual | 38.3937238 118 .325370541 R-squared = 0.1159

-------------+---------------------------------- Adj R-squared = 0.0485

Total | 43.427388 127 .341947937 Root MSE = .57041

------------------------------------------------------------------------------

uh | Coef. Std. Err. t P>|t| [95% Conf. Interval]

-------------+----------------------------------------------------------------

lchempi | -.1431572 .4720255 -0.30 0.762 -1.077896 .7915818

lgas | .6232994 .8859803 0.70 0.483 -1.131183 2.377782

lrtwex | .1786641 .3910344 0.46 0.649 -.5956904 .9530186

befile6 | -.0859232 .2510069 -0.34 0.733 -.5829851 .4111387

affile6 | -.1221207 .2546985 -0.48 0.632 -.6264931 .3822517

afdec6 | -.0668277 .2743671 -0.24 0.808 -.6101492 .4764937

uh |

L1. | .2214896 .0916573 2.42 0.017 .0399832 .4029959

L2. | .1340417 .0921595 1.45 0.148 -.0484592 .3165427

L3. | .125542 .0911194 1.38 0.171 -.0548992 .3059831

_cons | -14.36897 20.65581 -0.70 0.488 -55.27309 26.53516

where L1.uh L2.uh L3.uh are the first, second and third lags of the variable uhat.
3)
test L1.uh L2.uh L3.uh

( 1) L.uh = 0

( 2) L2.uh = 0

( 3) L3.uh = 0