SOLUTIONS TO SOME PROBLEMS:
1. Set of Problems 8, Problem 2: Consider the market of a product with a supply
function
Qsi = 0 + 1 Pi + usi ;
a demand function
Qdi = 0 + udi ;
and a market equilibrium condition, Qsi = Qdi ; where usi and udi are i.i.d. random
variable, mutually independent, both with zero mean.
(a) Show that Pi and usi are correlated.
.
We have to calculate Cov (Pi ; usi ) for which we need the reduced form of
Pi which is obtained solving the system of equations
Qi = 0 + 1 Pi + usi
Qi = 0 + udi ;
in terms of the exogenous variables and an error term, i.e. assuming 1 6= 0;
1
0 + 1 Pi + usi = 0 + udi ! Pi = 0 0 + udi usi
1
so that
1
Cov (Pi ; usi ) = Cov 0 0 + udi usi ; usi by the RF of Pi
1
1
= Cov udi usi ; usi by properties of Cov
1
1
= Cov (usi ; usi ) by independence of udi and usi
1
2
1 s
= V ar (usi ) = 6= 0:
1 1
1
(b) Show the OLS estimate of 1 is inconsistent.
The OLS of 1 is
d
^ = Cov (Qi ; Pi )
1
Vdar (Pi )
where from the RF Qi = 0 + udi we have
d (Qi ; Pi ) = Cov
d 1
Cov 0 + udi ; 0 0 + udi usi
1
1 d d d
= Cov ui ; ui usi
1
1
Cov udi ; udi usi
1
2
1 1 d
= Cov udi ; udi = V ar udi =
1 1 1
and from the RF of Pi ;
1
Vd
ar (Pi ) = Vd
ar 0 0 + udi usi
1
1 d d
= 2 V ar ui usi
1
1
2V ar udi usi
1
1
= 2 V ar udi + V ar (usi )
1
2 2
d + s
= 2 ;
1
so that 2
d 2
^ Cov (Qi ; Pi ) 1 d
1 = 2+ 2 = 1 2 2
6= 1
V ar (Pi ) d
2
s
d + s
1
2
con…rming that the OLS estimate is inconsistent unless s = 0:
(c) For estimation of 0; 1 we would need an instrument for the supply equa-
tion, i.e. a variable that is not correlated with usi but is correlated with
2
Pi : In the model the only candidate is Qi ; which satis…es
Cov Qi ; uSi = Cov 0 + udi ; uSi = Cov udi ; usi = 0
and
1
Cov (Qi ; Pi ) = Cov 0 + udi ; 0 0 + udi usi
1
1
= Cov udi ; udi usi
1
2
1 d
= V ar udi = 6= 0;
1 1
so that the IV of 1 satis…es for large n
d 2
^ IV = Cov (Qi ; Qi ) V ar (Qi )
= d
= 1:
1 2
d (Qi ; Pi )
Cov Cov (Qi ; Pi ) d
1
3
Set of Problems 9, Problem 2: Use the regressions in the following table to
answer the questions. For that data from 48 states of the USA were used.
The statistics were statistically signi…cant at the 5% or 1% level.
Dependent Variable: ln(Qcigarettes
i;1995 ) ln(Qcigarettes
i;1985 )
Regressor (1) (2) (3)
cigarretes cigarretes
ln(Pi;1995 ) ln(Pi;1985 ) 0:94 1:34 1:20
(0:21) (0:23) (0:20)
ln(Incomei;1995 ) ln(Incomei;1985 ) 0:53 0:43 0:05
(0:34) (0:30) (0:06)
Intercept 0:12 0:02 0:05
(0:07) (0:07) (0:06)
Speci…c tax Sales tax and Speci…c
Instrumental V ariables Sales Tax cigarettes tax cigarettes
F Statistic of the …rst stage 33.70 107.20 889.60
J T est f or restrictions
overidentif ication and p value
- - 4:93
(0:026)
(a) Imagine that the federal government is considering a new tax on the ciga-
rettes which is estimated will increase the retail price in 0.50$ per packet.
If the current price per packet is now 7.40$, use the regression in column (1)
to predict the variation in demand. Construct a 95% con…dence interval
for the variation in demand.
We take by granted that we are considering a panel regression model of the form
ln(Qcigarretes
i;t )= i + 1
cigarretes
ln(Pi;t )+ 2 ln(Incomei;t ) + uit ;
4
where i is a …xed e¤ect. Therefore,
ln(Qcigarretes
i;1995 ) ln(Qcigarretes
i;1985 ) = 0 + 1
cigarretes
ln(Pi;1995 ) cigarretes
ln(Pi;1985 )
+ 2 (ln(Incomei;1995 ) ln(Incomei;1985 )) + ui1995 ui1985 ;
where 0 = 0: Naturally, one can estimate the model with or without the intercept.
In this case the model has been estimated with intercept. However, we impose the
restriction that the intercept is equal zero.
cigarretes
1. (a) We …rst have to calculate the change in the regressor ln(Pi;1995 )
cigarretes
ln(Pi;1985 ) after the new tax with ln(Incomei;1995 ) ln(Incomei;1985 ) =
0:
cigarretes cigarretes
ln(Pi;1995 ) ln(Pi;1985 ) = ln(7:40 + 0:50) ln(7:40) = 0:0654
which is approximately equal to the relative change in prices due to the
new tax:
0:50
= 0:0676;
7:40
i.e. an increment of prices of 6.76%.
Then, the estimated change in demand is ^ 1 ln (P ) = 0:94 0:0654 =
0:0615; which implies a reduction of demand of 6.15%. The 95% con…-
dence interval is
0:0615 1:96SE ^ 1 0:0654 ! 0:0615 1:96SE ^ 1 0:0654
! 0:0615 1:96 0:21 0:0654
! ( 0:0884; 0:0346)
! ( 8:84%; 3:46%) :
Note the answer including the intercept, i.e.
cigarretes cigarretes
ln(Pi;1995 ) ln(Pi;1985 ) = ln(7:40 + 0:50) ln(7:40) = 0:0654
5
b. Imagine that the USA gets in recession and the income is reduced by 2%.
Use the regression in column (1) to predict the variation in demand.
In this case ln(Incomei;1995 ) ln(Incomei;1985 ) = 0:02; so that the e¤ect
on the demand is ^ 2 ln (Income) = 0:53 ( 0:02) = 0:010 6; i.e. a
reduction of 1.06% in the demand.
c. Imagine that the recession last less than one year. Do you think that the
regression in column (1) provides a reliable answer to the question in (b)?
Why or why not?
The regression in column (1) will not provide a reliable answer to the
question in (b) when recessions last less than 1 year. The regression in
column (1) studies the long-run price and income elasticity. Cigarettes are
addictive. The response of demand to an income decrease will be smaller
in the short run than in the long run.
Note that if the recession lasts less than one year and the previous level
of income is recovered before the year 1995, also the demand of cigarettes
would recover so that the regression in (1) can not exploit such short
run variations, only long run (10 years horizon) ones, which may have a
di¤erent impact as many other factors might adjust in between.
d. Imagine that the F statistics in column (1) were 3.6 instead of 33.70.
Would the regression provide a reliable answer to the question set in (a)?
Why or why not?
6
No, because a value F < 10 indicates the presence of weak instruments
(i.e. the …rst stage provides little information of the exogenous component
in the variation of the endogenous regressor), so that the IV estimate
could have a large bias and the normal distribution does not approximate
precisely the actual …nite sample properties of the estimate so we cannot
rely on the standard methods for statistical inference.
Set of Problems 9, Problem 6: The following model measures the e¤ect of
a programme of school choice in terms of the score of an standardized test,
score = 0 + 1 choice + 2 faminc + u1 ;
where score is the score of the test conducted at state level, choice is the
binary variable which indicates if a student has attended a choice school
during the last year, and faminc is the family income. The IV for choice
is grant; the amount in dollars received by the students to pay the fees of
the choice school. The grant received varies over the income level of the
family, which is the reason why we control for faminc in the equation.
(b) Even with faminc in the equation, why choice could be correlated with
u1 ?
In the error term u1 there are contained other factors which explain score
beyond choice and f aminc, e.g. those related to ability/intelligence, mo-
tivation, parents involvement in children education, etc., that might them-
selves being correlated to choice:
(c) If within each income level, the amount of the grant were randomly as-
signed, would grant be correlated with u1 ?
7
.
No, because this random assignment will make grant independent of u1
conditional on f aminc, i.e. grant is only correlated to f aminc, but u1 is
uncorrelated to f aminc (because f aminc is included in the model) and
therefore to grant:
(d) Write down the reduced form equation for choice: Which is needed for
grant to be partially correlated with choice?
The reduced form is
choice = 0 + 1 faminc + 2 grant +v
and for grant to be partially correlated with choice it is needed that 2 6= 0:
(e) Write down the reduced form equation for score: Explain why it could be
useful [Hint: how would you interpret the coe¢ cient of grant?].
The reduced form is
score = 0 + 1 faminc + 2 grant + w;
which is useful to interpret the coe¢ cient 1 as the ratio 2= 2; where 2 is
the average increment in score due a exogenous unitary increment in grant
(whose e¤ect is only transmitted to score through changes in choice; as
we assume that grant was randomly assigned, conditional on f aminc).
