Ec220 - Introduction to Econometrics

Practice Questions
1. (Adapted from S&W ch. 7 exercises) Consider the following regression results for av-
erage hourly earnings for American workers aged 25 to 34 in 1998. The variables in
the regression are the following:
College dummy variable, 1 = college degree, 0 = high school
Female dummy variable, 1 = female, 0 = male
Age age in years
Northeast dummy variable, 1 = lives in the Northeast, 0 = lives elsewhere
Midwest dummy variable, 1 = lives in the Midwest, 0 = lives elsewhere
South dummy variable, 1 = lives in the South, 0 = lives elsewhere
West dummy variable, 1 = lives in the West, 0 = lives elsewhere
The variables Northeast, Midwest, South, and West indicate one of four broad regions
of the US.
Regressor (1) (2) (3) (4)
College
5:46
(0:21)
5:48
(0:21)
5:44
(0:21)
A
(B)
Female
2:64
(0:20)
2:62
(0:20)
2:62
(0:20)
C
(D)
Age
0:29
(0:04)
0:29
(0:04)
E
(F)
Northeast
0:69
(0:30)
G
(H)
Midwest
0:60
(0:30)
I
(J)
South
0:27
(0:26)

West
K
(L)
Constant
12:69
(0:14)
4:40
(1:05)
3:75
(1:06)
M
(N)
(a) Explain why the coecients on College and Female change little between the
dierent specications in columns (1) to (3).
(b) Explain why the standard errors on College and Female change little between the
dierent specications in columns (1) to (3).
(c) What is the interpretation of the constant term (the number 12.69) in col. (1)?
(d) Column (3) includes regional dummy variables. Explain why the variable West
is omitted in this regression.
(e) The regression in column (4) omits the variable South and includes West. Fill in
as many of the coecients A to N as you can.
(f) What is the interpretation of the constant term (the number 3.75) in col. (3)?
1
(g) What are the expected average earnings of a 28 year old female college graduate
from the South?
(h) What is the dierence in earnings between a 26 year old male college graduate
from the Northeast and a 29 year old female high school graduate from the West?
(i) Howwould you specify the regression if you wanted to test whether wages of female
workers increase with age by the same amount as the wages of male workers?
(j) How are the residuals from the regression in col. (3) related to those from col.
(4)?
(k) How is the R
2
from the regression in col. (2) related to that from the regression
in col. (3)? How is the R
2
from the regression in col. (3) related to that from the
regression in col. (4)?
(l) How would you specify the regression to impose the restriction that the coecient
on Northeast and Midwest are the same?
(m) Explain whether the estimated coecient on College reects the causal eect of
graduating from college as compared to graduating from high school.
(n) Explain whether the estimated coecient on Female reects the eect of discrim-
ination against women in the labour market.
2. (Adapted from S&W exercise 12.9) We are interested in the eect of military service
on later civilian earnings. Suppose we have a sample of American men born in 1950
and we observe their earnings Y
i
in 1990, when they are 40 years old. A dummy
variable Serv
i
indicates whether the individual served during the Vietnam war. We
are interested in running the regression
ln (Y
i
) = +Serv
i
+e
i
:
(a) Explain whether you expect military service to have a causal eect on later civilian
earnings. What would be the causal channel. Do you think the eect is positive
or negative?
(b) What potential confounders can you think of? Which way would the bias from
these confounders go?
(c) During the Vietnam War there was a draft and priority for the draft was based
on a national lottery. The lottery consisted of assigning every birthday a number
from 1 (for 1 January) to 365 (for 31 December), and then ordering these numbers
randomly. Men whose birthdays came up at the top of the list were drafted rst,
then the next lower numbers, and so forth until the military lled all its needs for
the year (men born in 1950 were mostly drafted in 1970).
1. Are the draft lottery numbers a potentially plausible instrument for Vietnam
War service?
2. Many men volunteered for service during the Vietnam War. Volunteers
typically got better conditions of service, e.g. they might have had more
choice of units of the military to serve in compared to those called up through
the draft. As a result, more men with low lottery numbers (those being
more likely to be eventually drafted) volunteered than those with high lottery
numbers. Does this violate any of the IV assumptions, which one, and why?
2
3. A company producing laundry detergent is interested in the eect of its TV advertising
on sales. You have the following monthly data for 56 TV markets over 3 years
sales sales in TV market i in month t
price the price charged in TV market i in month t in constant 2000
competitors number of competitors selling detergent in market i in month t
ads the number of TV ads which ran in TV market i in month t
You dont have any other data on the customers in dierent TV markets.
(a) Suppose the rm allocates TV ads randomly to TV markets and months. What
regression would you run to estimate the causal eect of ads on sales? Write down
the regression equation and method of estimation, and explain why you include
the right hand side variables you do.
(b) Would your regression include the variable price on the RHS? Explain why or
why not.
(c) Continue to assume that ads are placed randomly. How would you specify a
regression model to test the hypothesis that TV advertising matters more for
sales in more competitive markets (those in which 3 or more competitors also see
laundry detergent). Be precise about how you would carry out the statistical test.
(d) Suppose that a regional marketing manager responsible for a particular TVmarket
decides on whether to run ads in their region.
1. Explain why a simple regression of sales on ads may not identify the causal
eect of ads. In which direction do you think any bias would go?
2. How would you go about estimating the causal eect of the TV ads now?
What additional assumptions do you need to make compared to the case in
a)?
(e) Suppose the rm runs ads only in markets when two or more competitors are
present. How would you go about estimating the causal eect of the TV ads
now?
(f) Suppose you have data on the following additional variables:
i) ads run by competitors in the same market and month,
ii) sales in the region in the previous month.
Would these variables be useful as instruments to identify the causal eect of TV
ads on sales?
3
4. This question refers to the HMDA data set on mortgage applicants, which we have used in the
lecture on binary choice models. The following variables are used below:

deny a dummy variable for whether an applicant was denied the mortgage (mean =
0.120)
single a dummy variable for whether the applicant is single
condo a dummy variable, 1 = the applicant is buying a condominium (flat), 0 = the
applicant is buying a house
sin_con the product single*condo
pi_rat the total monthly debt repayments divided by the monthly income of the household
ccred the consumer credit score; a higher score means a worse credit rating
school years of education of the applicant

a) The following is Stata output from a linear probability model and a probit on the single dummy:

. reg deny single, robust

Linear regression Number of obs = 2357
F( 1, 2355) = 13.64
Prob > F = 0.0002
R-squared = 0.0062
Root MSE = .32466

------------------------------------------------------------------------------
| Robust
deny | Coef. Std. Err. t P>|t| [95% Conf. Interval]
-------------+----------------------------------------------------------------
single | .0525721 .0142343 3.69 0.000 .024659 .0804852
_cons | .0998603 .0079262 12.60 0.000 .0843173 .1154034
------------------------------------------------------------------------------

. probit deny single, robust

Probit regression Number of obs = 2357
Wald chi2(1) = 14.41
Prob > chi2 = 0.0001
Log pseudolikelihood = -859.96661 Pseudo R2 = 0.0083

------------------------------------------------------------------------------
| Robust
deny | Coef. Std. Err. z P>|z| [95% Conf. Interval]
-------------+----------------------------------------------------------------
single | .2562911 .0675203 3.80 0.000 .1239537 .3886285
_cons | -1.282348 .0452004 -28.37 0.000 -1.370939 -1.193757
------------------------------------------------------------------------------

Explain formally how the coefficients on female in the probit and the linear probability models are
related.
Consider the following regression output:

. reg deny single pi_rat ccred, robust

Linear regression Number of obs = 2357
F( 3, 2353) = 54.78
Prob > F = 0.0000
R-squared = 0.1089
Root MSE = .30756

------------------------------------------------------------------------------
| Robust
deny | Coef. Std. Err. t P>|t| [95% Conf. Interval]
-------------+----------------------------------------------------------------
single | .0455956 .0135127 3.37 0.001 .0190976 .0720936
pi_rat | .5638007 .0868146 6.49 0.000 .3935595 .7340418
ccred | .0493021 .0050359 9.79 0.000 .0394269 .0591774
_cons | -.188203 .0285917 -6.58 0.000 -.2442705 -.1321354
------------------------------------------------------------------------------

. reg deny single pi_rat ccred school, robust

Linear regression Number of obs = 2357
F( 4, 2352) = 42.17
Prob > F = 0.0000
R-squared = 0.1129
Root MSE = .30693

------------------------------------------------------------------------------
| Robust
deny | Coef. Std. Err. t P>|t| [95% Conf. Interval]
-------------+----------------------------------------------------------------
single | .0460667 .0134854 3.42 0.001 .0196221 .0725113
pi_rat | .5545216 .0909171 6.10 0.000 .3762356 .7328077
ccred | .048763 .0050368 9.68 0.000 .0388859 .0586402
school | -.0072373 .0021847 -3.31 0.001 -.0115214 -.0029531
_cons | -.0720306 .0456616 -1.58 0.115 -.1615718 .0175106
------------------------------------------------------------------------------

b) The coefficient on school in the second regression seems important while the coefficient on
single changes little when school is added. Explain why this is the case.

c) Explain why the standard error on pi_rat has increased in the second regression while the
standard error on ccred has basically stayed the same.

Consider the following regression output:

. reg deny single sin_con condo pi_rat ccred , robust

Linear regression Number of obs = 2357
F( 5, 2351) = 34.18
Prob > F = 0.0000
R-squared = 0.1118
Root MSE = .3072

------------------------------------------------------------------------------
| Robust
deny | Coef. Std. Err. t P>|t| [95% Conf. Interval]
-------------+----------------------------------------------------------------
single | .06748 .0176357 3.83 0.000 .0328968 .1020631
sin_con | -.0788836 .0315525 -2.50 0.012 -.1407573 -.0170099
condo | .0503891 .0222421 2.27 0.024 .0067729 .0940053
pi_rat | .5667692 .0867344 6.53 0.000 .3966854 .7368531
ccred | .0489473 .0050344 9.72 0.000 .0390751 .0588196
_cons | -.1974432 .0288937 -6.83 0.000 -.254103 -.1407834
------------------------------------------------------------------------------

d) What is the effect of being single compared to a married couple with similar characteristics
buying a house on the probability of being denied a mortgage?

e) What is the effect of being single compared to a married couple with similar characteristics
buying a condo on the probability of being denied a mortgage?

f) At what significance level is the interaction term sin_con significantly different from zero?

g) How would you carry out a test of whether applicants buying condos are more likely to be denied
mortgages than house buyers?

h) Based on the various regression results above, do you think there is a causal effect of being
single on being denied a mortgage? How large is this effect? Which additional results would you
like to see to assess this question?


5. The following Stata output is produced using the data set on lecture attendance from problem set
7.

. summarize attend fr_att frosh priGPA ACT

Variable | Obs Mean Std. Dev. Min Max
-------------+--------------------------------------------------------
attend | 680 26.14706 5.455037 2 32
fr_att | 680 6.029412 11.25743 0 32
frosh | 680 .2323529 .4226438 0 1
priGPA | 680 2.586775 .5447141 .857 3.93
ACT | 680 22.51029 3.490768 13 32

. reg stndfnl attend fr_att frosh priGPA ACT, robust

Linear regression Number of obs = 680
F( 5, 674) = 33.95
Prob > F = 0.0000
R-squared = 0.2044
Root MSE = .88583

------------------------------------------------------------------------------
| Robust
stndfnl | Coef. Std. Err. t P>|t| [95% Conf. Interval]
-------------+----------------------------------------------------------------
attend | .0204009 .0081117 2.51 0.012 .0044736 .0363281
fr_att | -.021336 .0140661 -1.52 0.130 -.0489548 .0062827
frosh | .6224787 .3719511 1.67 0.095 -.1078435 1.352801
priGPA | .4234329 .086018 4.92 0.000 .2545375 .5923284
ACT | .0832826 .0110908 7.51 0.000 .061506 .1050592
_cons | -3.489796 .306119 -11.40 0.000 -4.090858 -2.888735
------------------------------------------------------------------------------

The variable stndfnl is the standardised final exam mark in the course (i.e. a variable with mean zero
and std. deviation one), attend is the number of lectures attended, frosh is a dummy variable for first
year students, fr_att is the interaction frosh*attend, priGPA is prior grade point average, and
ACT is a score from a test taken before entering university.

a) Can you reject the hypothesis that a one unit higher ACT score raises final exam marks by 0.10
standard deviations at the 5% level? Can you reject the hypothesis at the 1% level? Can you
reject the hypothesis at the 10% level?
b) What is the interpretation of the coefficient on frosh?

c) What is the interpretation of the coefficient on attend?

d) A commentator suggests that the regression should include a variable measuring the number of
missed lectures. Explain whether you agree or disagree with this comment.

e) By how much do you expect the final exam result of a first year student to differ if she attends five
more lectures?

f) How different are the final exam results of a first year and a later year student attending the
average number of lectures?

g) Do you think the effect of being a first year student can be interpreted causally in the regression
above? J ustify your answer.









6. You have data from 1959 to 2009 on the following two variables:
temp: relative global annual average temperature in 0.01 degrees Celsius, base 1950-81
co2: annual average atmospheric carbon dioxide concentration in ppm., Mauna Loa,
Hawaii



and you are presented with the following Stata results:

. reg temp co2, robust

Linear regression Number of obs = 51
F( 1, 49) = 218.58
Prob > F = 0.0000
R-squared = 0.8026
Root MSE = 9.9499

------------------------------------------------------------------------------
| Robust
temp | Coef. Std. Err. t P>|t| [95% Conf. Interval]
-------------+----------------------------------------------------------------
co2 | .9190583 .0621644 14.78 0.000 .7941342 1.043982
_cons | -298.9318 21.77594 -13.73 0.000 -342.6921 -255.1714
------------------------------------------------------------------------------

. reg temp co2 year, robust

Linear regression Number of obs = 51
F( 2, 48) = 118.78
Prob > F = 0.0000
R-squared = 0.8318
Root MSE = 9.2797

------------------------------------------------------------------------------
| Robust
temp | Coef. Std. Err. t P>|t| [95% Conf. Interval]
-------------+----------------------------------------------------------------
co2 | 2.462617 .5214701 4.72 0.000 1.414131 3.511103
year | -2.258271 .7422222 -3.04 0.004 -3.750609 -.7659334
_cons | 3646.721 1293.088 2.82 0.007 1046.794 6246.649
------------------------------------------------------------------------------


. dfuller temp

Dickey-Fuller test for unit root Number of obs = 50

---------- Interpolated Dickey-Fuller ---------
Test 1% Critical 5% Critical 10% Critical
Statistic Value Value Value
------------------------------------------------------------------------------
Z(t) -1.828 -3.580 -2.930 -2.600
------------------------------------------------------------------------------
MacKinnon approximate p-value for Z(t) = 0.3665
3
2
0
3
4
0
3
6
0
3
8
0
4
0
0
C
O
2

c
o
n
c
.
,

p
p
m
-
.
2
0
.
2
.
4
.
6
t
e
m
p
e
r
a
t
u
r
e
,

d
e
g

C
e
l
s
i
s
u
s
1960 1970 1980 1990 2000 2010
year
temperature, deg Celsisus CO2 conc., ppm
. dfuller co2

Dickey-Fuller test for unit root Number of obs = 50

---------- Interpolated Dickey-Fuller ---------
Test 1% Critical 5% Critical 10% Critical
Statistic Value Value Value
------------------------------------------------------------------------------
Z(t) 5.445 -3.580 -2.930 -2.600
------------------------------------------------------------------------------
MacKinnon approximate p-value for Z(t) = 1.0000

. dfuller temp, trend

Dickey-Fuller test for unit root Number of obs = 50

---------- Interpolated Dickey-Fuller ---------
Test 1% Critical 5% Critical 10% Critical
Statistic Value Value Value
------------------------------------------------------------------------------
Z(t) -5.399 -4.150 -3.500 -3.180
------------------------------------------------------------------------------
MacKinnon approximate p-value for Z(t) = 0.0000

. dfuller co2, trend

Dickey-Fuller test for unit root Number of obs = 50

---------- Interpolated Dickey-Fuller ---------
Test 1% Critical 5% Critical 10% Critical
Statistic Value Value Value
------------------------------------------------------------------------------
Z(t) -1.311 -4.150 -3.500 -3.180
------------------------------------------------------------------------------
MacKinnon approximate p-value for Z(t) = 0.8852

. reg d.temp d.co2, robust

Linear regression Number of obs = 50
F( 1, 48) = 8.56
Prob > F = 0.0052
R-squared = 0.1276
Root MSE = 12.313

------------------------------------------------------------------------------
| Robust
D.temp | Coef. Std. Err. t P>|t| [95% Conf. Interval]
-------------+----------------------------------------------------------------
co2 |
D1. | 7.977158 2.726606 2.93 0.005 2.494949 13.45937
_cons | -10.36659 4.477012 -2.32 0.025 -19.36823 -1.364959
------------------------------------------------------------------------------


. reg d.temp d.co2 year, robust

Linear regression Number of obs = 50
F( 2, 47) = 6.34
Prob > F = 0.0036
R-squared = 0.1719
Root MSE = 12.123

------------------------------------------------------------------------------
| Robust
D.temp | Coef. Std. Err. t P>|t| [95% Conf. Interval]
-------------+----------------------------------------------------------------
co2 |
D1. | 11.87479 3.377353 3.52 0.001 5.080436 18.66915
year | -.2447391 .1439168 -1.70 0.096 -.5342624 .0447841
_cons | 469.7547 282.3302 1.66 0.103 -98.22035 1037.73
------------------------------------------------------------------------------

Comment on these results. What, if anything, do you conclude about the impact of increasing
carbon dioxide concentrations on the global temperature, and why? What other statistical
evidence would you like to see to address this problem? Give some specific examples of what
you would do.