Homework Assignment 4

Carlos M. Carvalho
McCombs School of Business

Problem 1
Suppose we are modeling house price as depending on house size, the number of bedrooms
in the house and the number of bathrooms in the house. Price is measured in thousands of
dollars and size is measured in thousands of square feet.
Suppose our model is:

P = 20 + 50 size + 10 nbed + 15 nbath + ,  ∼ N (0, 102 ).

(a) Suppose you know that a house has size =1.6, nbed = 3, and nbath =2.

What is the distribution of its price given the values for size, nbed, and nbath.
(hint: it is normal with mean = ?? and variance = ??)
20 + 50 × 1.6 + 10 × 3 + 15 × 2 = 160
P = 160 +  so that P ∼ N (160, 102 )

(b) Given the values for the explanatory variables from part (a), give the 95% predictive
interval for the price of the house.
160 ± 20

(c) Suppose you know that a house has size =2.6, nbed = 4, and nbath =3. Give the
95% predictive interval for the price of the house.
20 + 50 × 2.6 + 10 × 4 + 15 × 3 = 235
P = 235 +  so that P ∼ N (235, 102 ) and the 95% predictive interval is
235 ± 20

(d) In our model the slope for the variable nbath is 15. What are the units of this
number?
Thousands of dollars per bathroom.

1
 (e) What are the units of the intercept 20? What are the units of the the error standard
deviation 10?
The intercept has the same units as P ... in this case, thousands of dollars. The error std
deviation is also in the same units as P , ie, thousands of dollars.

2
Problem 2
For this problem us the data is the file Profits.csv.
There are 18 observations.
Each observation corresponds to a project developed by a firm.
y = Profit: profit on the project in thousands of dollars.
x1= RD: expenditure on research and development for the project in thousands of dollars.
x2=Risk: a measure of risk assigned to the project at the outset.

We want to see how profit on a project relates to research and development expenditure
and “risk”.

(a) Plot profit vs. each of the two x variables. That is, do two plots y vs. x1 and y
vs x2. You can’t really understand the full three-dimensional relationship from these
two plots, but it is still a good idea to look at them. Does it seem like the y is related
to the x’s?

(b) Suppose a project has risk=7 and research and development = 76. Give the 95%
plug-in predictive interval for the profit on the project. Compare that to the correct,
predictive interval (using the predict function in R).

(c) Suppose all you knew was risk=7. Run the simple linear regression of profit on risk
and get the 68% plug-in predictive interval for profit.

(d) How does the size of your interval in (c) compare with the size of your interval in (b)?
What does this tell us about our variables?

(a) It seems like there is some relationship, especially between RD and profit.

(b) The plug-in predictive interval, when RD = 76 and RISK = 7 is 94.75 ± 2 ∗ 14.34 =
[66.1, 123.4].

(c) Using the model P ROF IT = β0 + β1 RISK + , the 68% plug-in prediction interval for
when RISK = 7 is 143 ± 106.1 = [37.5, 249.7].

3
(d) Our interval in (c) is bigger than the interval in (b) despite the fact that it is a “weaker”
confidence interval. In essence (c) says that we predict Y will be in [38, 250] 68% of
the time when RISK = 7. In contrast, (b) says that Y will be in [63, 127] 95% of
the time when RISK = 7 and RD = 76. Using RD in our regression narrows our
prediction interval by quite a bit.

4
Problem 3
The data for this question is in the file zagat.xls . The data is from the Zagat restaurant
guide. There are 114 observations and each observation corresponds to a restaurant.
There are 4 variables:
price: the price of a typical meal
food: the zagat rating for the quality of food.
service: the zagat rating for the quality of service.
decor: the zagat rating for the quality of the decor.

We want to see how the price of a meal relates the quality characteristics of the restaurant
experience as measured by the variables food, service, and decor.

(a) Plot price vs. each of the three x’s. Does it seem like our y (price) is related to the
x’s (food, service, and decor) ?

(b) Suppose a restaurant has food = 18, service=14, and decor=16. Run the regression
of price on food, decor, and service and give the 95% predictive interval for the price
of a meal.

(c) What is the interpretation of the coefficient estimate for the explanatory variable food
in the multiple regression from part (b) ?

(d) Suppose you were to regress price on the one variable food in a simple linear regression?
What would be the interpretation of the slope? Plot food vs. service. Is there a
relationship? Does it make sense? What is your prediction for how the estimated
coefficient for the variable food in the regression of price on food will compare to the
estimated coefficient for food in the regression of price on food, service, and decor?
Run the simple linear regression of price on food and see if you are right! Why are
the coefficients different in the two regressions?

(e) Suppose I asked you to use the multiple regression results to predict the price of a
meal at a restaurant with food = 20, service = 3, and decor =17. How would you feel
about it?

5
Does it seem like our y (price) is related to the x’s (food, service, and decor) ?

! ! !!
! !

60

60

60
! !
! ! ! ! !!!!
!
! ! ! ! ! !
50 ! ! !

50

50
! ! !
! !
! ! ! ! !
! ! ! ! ! !
!
! ! ! !
!! !! !!
! ! ! ! !
zd$price

zd$price

zd$price
! !
! ! !
! ! ! ! ! !!
! ! ! ! !
! ! ! !! ! ! !!
40

40

40
! ! ! !
! ! ! !!!
! !
! ! !
! ! ! ! !
! ! ! ! ! ! ! ! ! ! !! ! !!
! ! ! ! !
! !
! ! ! ! ! ! !! !!!
! ! ! !
! ! !
! ! !
! !
! ! !
! !! !
30

30

30
! ! ! ! ! !
! ! !
! ! ! !
! ! ! ! ! ! !
! ! ! ! ! ! !!! ! !!
! ! ! !
! !
! !! !!
! ! ! ! !
! ! ! ! ! ! ! ! ! ! !!!!
! ! ! ! ! !
!!!!!! !!
!
!
! !
! ! ! ! ! ! ! !
! ! !!
20

20

20
! ! ! ! ! ! ! ! !!
! ! ! ! ! !
!
! ! ! !
!!
! ! ! ! !
! ! ! ! ! !
10

10

10
14 16 18 20 22 24 26 10 15 20 25 5 10 15 20 25

zd$food zd$service zd$decor

Solutions.
Definitely looks like price is related to each of our 3 x’s.
(a) Check out the figure above... definitely looks like price is related to each of the 3 X’s.

(b) The regression output is

(b) SUMMARY OUTPUT

Regression Statistics
Multiple R 0.829
R Square 0.687
Adjusted R 0.679
Suppose a restaurant has food = 18, service=14, and decor=16.
Standard E 6.298
Observatio 114.000

ANOVA
df SS MS F Significance F
Run the regression
Regression
Residual
3.000 of price
110.000
on food,
9598.887 3199.629
4363.745 39.670
80.655 decor, and service and give the 95% plug-in predict
0.000

intervalTotalfor the113.000
price of a meal.
13962.632

Coefficients Standard Error t Stat P-value Lower 95% Upper 95%

Intercept -30.664 4.787 -6.405 0.000 -40.151 -21.177
food 1.380 0.353 3.904 0.000 0.679 2.080
decor 1.104 0.176 6.272 0.000 0.755 1.453
service 1.048 0.381 2.750 0.007 0.293 1.803
Estimate Std. Error t value Pr(>|t|)
(Intercept) -30.6640 4.7872 -6.405 3.82e-09 ***
food 1.3795 0.3533 3.904 0.000163 ***
so that −30.66 + 1.38 × 18 + 1.1 × 16 + 1.05 × 14 = 26.476 and the 95% plug-in
decor 1.1043
prediction 0.1761
interval is 26.476 ± 12.6 6.272 7.18e-09 ***
service(c) If you hold
1.0480 0.3811
service and decor 2.750
constant and increase0.006969 **price goes up (on
food by 1, then
--- average) by 1.38.
Signif.(d)codes:
If food goes0up*** 0.001
by 1 price goes** 0.01
up by * 0.05
the slope . 0.1 from1the plot in item
(on average)...
(a) we know that it looks like food and price are related in a positive way. Now, you
would think that these four variables are somewhat related to each other, right? A
better restaurant tend to have good food, service6 and decor... and also a higher price.
By running the regression with only food as a explanatory variable I would guess the
coefficient for food would be higher... let’s see:

6
SUMMARY OUTPUT

Regression Statistics
Multiple R 0.599
R Square 0.359
Adjusted R Squ 0.353
Standard Error 8.939
Observations 114.000

ANOVA
df SS MS F Significance F
Regression 1.000 5012.239 5012.239 62.720 0.000
Residual 112.000 8950.393 79.914
Total 113.000 13962.632

Coefficients Standard Error t Stat P-value Lower 95% Upper 95%

Intercept -18.154 6.553 -2.770 0.007 -31.137 -5.170
food 2.625 0.331 7.920 0.000 1.968 3.282

I was right! In the simple linear, regression food works as a proxy for the overall
quality of a restaurant. When food goes up service and decor tend to go up as well
but since they are not in the regression, the coefficient for food has to reflect the other
factors. Once decor and service are in the regression, the coefficient for food just has
to reflect the impact associated with food but not with the other variables.

(e) Very bad! We just dont see in our data restaurants with that low of a service rating
given food equal to 20 and decor equal to 17. This would be a extreme extrapolation
from what we have seen so far and the model might not be appropriate.

7
Problem 4: Baseball
Using our baseball data (RunsPerGame.xls), regress R/G on a binary variable for league
membership (League = 0 if National and League = 1 if American) and OBP .

R/G = β0 + β1 League + β2 OBP + 

1. Based on the model assumptions, what is the expected value of R/G given OBP for
teams in the AL? How about the NL?
2. Interpret β0 , β1 and β2 .
3. After running the regression and obtaining the results, can you conclude with 95%
probability that the marginal effect of OBP on R/G (after taking into account the
League effect) is positive?
4. Test the hypothesis that β1 = 0 (with 99% probability). What do you conclude?

1. The expected value of R/G given OBP is

h i
E R/G|OBP, League = 0 = β0 + β2 OBP

for the NL and

E[R/G|OBP, League = 1] = (β0 + β1 ) + β2 OBP

for the AL.

2. β0 is the number of runs per game we expect a team from the National League to
score if their OBP is zero.
We expect a team in the American League to score β1 more runs per game on average
than a team in the National League with the same OBP .
β2 tells us how R/G scales with OBP . For every unit increase in OBP there will be
a β2 increase in R/G.
3. The 95% confidence interval for β2 is 37.26±2*2.72 = (31.82;42.70) hence, yes, it is
greater than zero.
4. The best guess of β1 is b1 = 0.01615 with standard error 0.06560. Thus the 99%
confidence interval is b1 ± 3 ∗ sb1 = [−0.18, 0.21], which includes zero. Since zero is in
our interval of reasonable values we cannot conclude that β1 6= 0.
Coefficients:
Estimate Std. Error t value Pr(>|t|)
(Intercept) -7.72065 0.93031 -8.299 6.59e-09 ***
LeagueAmerican 0.01615 0.06560 0.246 0.807
OBP 37.26060 2.72081 13.695 1.14e-13 ***
---
Signif. codes: 0 *** 0.001 ** 0.01 * 0.05 . 0.1 1

Residual standard error: 0.1712 on 27 degrees of freedom

Multiple R-squared: 0.8851, Adjusted R-squared: 0.8765
F-statistic: 103.9 on 2 and 27 DF, p-value: 2.073e-13

8
Problem 5
Read the case “Orion Bus Industries: Contract Bidding Strategy” in the course packet.
Orion Bus Industries wants to develop a method for determining how to bid on specific bus
contracts to maximize expected profits. In order to do this, it needs to develop a model of
winning bids that takes into account such factors as the number of buses in the contract,
the estimated cost of the buses and the type of bus (e.g. length, type of fuel used, etc.).
The data set is available in the course website. This data set only includes the bus contracts
from Exhibit 1 in the case where Orion did not win the contract. This eliminates 28 of the
69 observations and leaves a sample of size n = 41 observations.

(a) Run a regression of W inningBid against N umberOf BusesInContract, OrionsEstimatedCost,

Length, Diesel and HighF loor, ie, the following regression model:

W inningBidi = β0 + β1 N umberOf BusesInContracti + β2 OrionsEstimatedCosti +

β3 Lengthi + β4 Dieseli + β5 HighF loori + i

What is the estimated regression model? How would you interpret the estimated
coefficient associated with the dummy variable Diesel?

(b) What is the estimate of σ 2 in the model in part (a)?

The city of Louisville, Kentucky is putting out a contract for bid for five 30-foot, low-floor,
diesel-fuelled buses. Orion estimates their cost to manufacture these buses to be $234,229
per bus.
(c) Using the model in part (a), what is the distribution representing the uncertainty
about the amount of the winning bid per bus for this contract? In particular, what
are the mean and standard deviation of the distribution?

(d) Given the distribution in part (c), what is the probability that Orion wins the contract
if it bids $240,000 per bus? If it wins the contract, what is its profit per bus per bus?

(e) What is the probability that Orion loses the contract if it bids $240,000 per bus? If it
loses the contract, what is its profit per bus? (You do not need to take into account
the cost of putting the bid together when determining the profit for a lost contract.)

(f) Why is there uncertainty about the profit per bus that Orion will obtain if it bids
$240,000 per bus? What is the probability distribution representing this uncertainty?
In particular, what is the mean of the distribution (i.e. what is the expected profit
per bus if it bids $240,000 per bus)?

We now want to develop an Excel spreadsheet that will allow ExpectedProfit to be plotted
against different possible bid amounts (i.e. $240,000; $241,000; ...; $260,000). The maximum
of this graph will give Orion the bid amount that will maximize expected profit.

(g) Using the plot, what should Orion bid if it wants to maximize expected profit per
bus?

9
The Excel regression output for this regression model is:
(a) The Excel output for this regression model is:
SUMMARY OUTPUT

Regression Statistics
Multiple R 0.902304784
R Square 0.814153923
Adjusted R Square 0.787604483
Standard Error 11721.15707
Observations 41

ANOVA
df SS MS
Regression 5 21065032698 4213006540
Residual 35 4808493307 137385523.1
Total 40 25873526005

Coefficients Standard Error t Stat

Intercept -13872.72734 26062.71483 -0.532282513
NumberOfBusesInContract 42.32044997 219.3318353 0.192951698
OrionsEstimatedCost 0.813616165 0.073356177 11.09131092
Length 1949.968943 456.482292 4.271729652
Diesel 11240.97951 6172.434639 1.821158128
HighFloor 8175.562414 4353.019803 1.878135819

Therefore, the estimated regression equation is

WinningBidi = -13872.7 + 42.3204 NumberOfBusesInContracti

The interpretation of the estimated coefficient for Diesel is the following:
+ 0.813616OrionsEstimatedCosti + 1949.97Lengthi + 11241.0 Dieseli

First,+ 8175.56
the true coefficient β4 is the expected increase, on average, in the winning bid
HighFloor i
when the buses specified in the contract run on diesel fuel rather than natural gas,
The interpretation
holdingofall
theother
estimated coefficient
variables associated with Diesel is the following:
constant.
First, the true coefficient E4 is the expected increase, on average, in the winning bid
when the11241.0
buses specified
is thein the contract run
estimate for on
β4diesel
(ie bfuel rather than natural gas, holding
4 in our notation) so that 11241.0 is the estimate
all other variables constant.
of the expected increase, on average, in the winning bid when the buses specified in
theis contract
11241.0 the estimaterun
of E4 on diesel isfuel
so 11241.0 ratherofthan
the estimate natural
the expected gas, on
increase, holding all other variables
average, in the winning bid when the buses specified in the contract run on diesel fuel
constant.
rather than natural gas, holding all other variables constant.
(b) s = 11721.15

(c) For a contract with five 30-foot, low-floor, diesel-fuelled buses and an estimated cost
of $234,229 per bus, the explanatory variables take on the following values:
N umberOf BusesInContract
7 = 5; OrionsEstimatedCost = 234, 229; Length = 30;
Diesel = 1 and HighF loor = 0.

Given the estimates from (a), the estimated mean of the distribution is
-13872.7 + 42.3204(5) + 0.813616(234229) + 1949.97(30) + 11241.0(1) + 8175.56(0)
= 246651.5.
so that the distribution of the winning bid can be represented by

W inningBid ∼ N (246651.5, 117212 )

10
 (d) To find the probability that Orion wins the contract if it bids $240,000 per bus we need
to compute the following probability (note that LowBid is the same as WinningBid
but is a bit more descriptive of what the above regression provides):

P r(Win Contract) = P r(Low Bid > 240000) = 0.7146

If Orion wins the contract, Profit (which is the difference between the bid amount of
$240,000 and the cost of $234,229) is $5,771.

(e) The probability that Orion loses the contract is

P r(Lose Contract) = 1 − P r(Win Contract) = 0.2854

If Orion loses the contract, then it receives no revenue and has no production costs
so its Profit is 0.

(f) There is uncertainty about the profit that Orion will obtain because there is uncer-
tainty about whether the company will win the contract or not. The probability
distribution representing the uncertainty is

Profit Probability
$0 0.2854
$5,771 0.7146
(h) To plot ExpectedProfit (column D) against BidAmount (column A), copy column A to
column F, and
Thiscolumn D to column
distribution G (BidAmount
has a mean ofand ExpectedProfit must be in
adjacent columns for plotting) and then use the Chart Wizard to create the plot. When
you copy column D it is important to use the command Edit > Paste Special > Values to
Expected Profit = E(P rof it) = $0(0.2854) + $5771(0.7146)
paste the values into column G (because column D depends on a formula).
= $4124

Using (g)
the plot,
Thewhat
plotshould Orion bid ifProfit
of Expected it wantsversus
to maximize expected profit?
Bid Amount is
The plot of ExpectedProfit against BidAmount is

ExpectedProfit vs. BidAmount

7000

6000

5000

4000
3000

2000

1000

0
235000 240000 245000 250000 255000 260000 265000

The maximum ExpectedProfit in the graph occurs at approximately $248,000. Therefore,

Orion should bid $248,000 per bus.
The maximum Expected Profit in the graph occurs at approximately $248,000. There-
fore, Orion should bid $248,000 per bus.

11
Problem 6: Beauty Pays!
Professor Daniel Hamermesh from UT’s economics department has been studying the im-
pact of beauty in labor income (yes, this is serious research!!).

First, watch the following video:

http://www.thedailyshow.com/watch/mon-november-14-2011/ugly-people

It turns out this is indeed serious research and Dr. Hamermesh has demonstrated the effect
of beauty into income in a variety of different situations. Here’s an example: in the paper
“Beauty in the Classroom” they showed that “...instructors who are viewed as better looking
receive higher instructional ratings” leading to a direct impact in the salaries in the long
run.

By now, you should know that this is a hard effect to measure. Not only one has to work
hard to figure out a way to measure “beauty” objectively (well, the video said it all!) but
one also needs to “adjust for many other determinants” (gender, lower division class, native
language, tenure track status).

So, Dr. Hamermesh was kind enough to share the data for this paper with us. It is available
in our class website in the file “BeautyData.csv”. In the file you will find, for a number
of UT classes, course ratings, a relative measure of beauty for the instructors, and other
potentially relevant variables.

1. Using the data, estimate the effect of “beauty” into course ratings. Make sure to
think about the potential many “other determinants”. Describe your analysis and
your conclusions.
We talked about this one in class. The main point here is that in order to isolate the
effect of beauty into class ratings we need to CONTROL for other potential deter-
minants of ratings. From the data available it looks like all the other variables are
relevant so we should be running the following regression:

Ratings = β0 + β1 BeautyScore + β2 F emale + β3 Lower + β4 N onEnglish +

β5 T enureT rack + 

12
 Here are the results:

Coefficients:
Estimate Std. Error t value Pr(>|t|)
(Intercept) 4.06542 0.05145 79.020 < 2e-16 ***
BeautyScore 0.30415 0.02543 11.959 < 2e-16 ***
female -0.33199 0.04075 -8.146 3.62e-15 ***
lower -0.34255 0.04282 -7.999 1.04e-14 ***
nonenglish -0.25808 0.08478 -3.044 0.00247 **
tenuretrack -0.09945 0.04888 -2.035 0.04245 *

So, as discussed in class it makes sense for some of these coefficients to be negative,
right? For example, if an instructor is not a native english speaker he/she might have a
harder time communicating the material and hence lower teaching evaluations. Same
goes for lower division classes; most people have to take those classes whether they
want or not which leads to lower ratings as students are potentially less interested in
the materials to begin with. Now, the results for females is a bit surprising. Why
are (holding all else equal) females instructors receiving lower ratings on average?
Are there any reasons for us to believe females are not as capable as males to teach?
Probably not, right? So, this data demonstrates a potential negative bias that people
have in evaluating women.
Finally, with all of that taken into account we find that the higher the beauty score
of the instructor the higher their ratings!

2. In his paper, Dr. Hamermesh has the following sentence: “Disentangling whether
this outcome represents productivity or discrimination is, as with the issue generally,
probably impossible”. Using the concepts we have talked about so far, what does he
mean by that?
The question here is: are beautiful people indeed better teachers or are they just
perceived to be better teachers because of their looks? This analysis can’t answer this
question! In my opinion the results are very suggestive that this is just discrimination
as I dont really believe that beauty relates to one’s ability to teach. But, until we run
an controlled experiment or find a “natural experiment” (like the one in question 3)
we can’t conclusively prove this point. What would be a potential natural experiment
here? Wouldn’t it be nice if we had data on blind students taking these classes? Why
would that help?

13
Problem 7: Housing Price Structure
The file MidCity.xls, available on the class website, contains data on 128 recent sales of
houses in a town. For each sale, the file shows the neighborhood in which the house is
located, the number of offers made on the house, the square footage, whether the house
is made out of brick, the number of bathrooms, the number of bedrooms, and the selling
price. Neighborhoods 1 and 2 are more traditional whereas 3 is a more modern, newer and
more prestigious part of town. Use regression models to estimate the pricing structure of
houses in this town. Consider, in particular, the following questions and be specific in your
answers:

1. Is there a premium for brick houses everything else being equal?

2. Is there a premium for houses in neighborhood 3?
3. Is there an extra premium for brick houses in neighborhood 3?
4. For the purposes of prediction could you combine the neighborhoods 1 and 2 into a
single “older” neighborhood?

There may be more than one way to answer these questions.

(1) To begin we create dummy variable Brick to indicate if a house is made of brick and
N2 and N3 to indicate if a house came from neighborhood two and neighborhood
three respectively. Using these dummy variables and the other covariates, we ran a
regression for the model
Y = β0 + β1 Brick + β2 N2 + β3 N3 + β4 Bids
+ β5 SqF t + β6 Bed + β7 Bath + ,  ∼ N (0, σ 2 ).
and got the following regression output.
Coefficients:
Estimate Std. Error t value Pr(>|t|)
(Intercept) 2159.498 8877.810 0.243 0.808230
BrickYes 17297.350 1981.616 8.729 1.78e-14 ***
N2 -1560.579 2396.765 -0.651 0.516215 2.5 % 97.5 %
N3 20681.037 3148.954 6.568 1.38e-09 *** (Intercept) -15417.94711 19736.94349
Offers -8267.488 1084.777 -7.621 6.47e-12 *** BrickYes 13373.88702 21220.81203
SqFt 52.994 5.734 9.242 1.10e-15 *** N2 -6306.00785 3184.84961
Bedrooms 4246.794 1597.911 2.658 0.008939 ** N3 14446.32799 26915.74671
Bathrooms 7883.278 2117.035 3.724 0.000300 *** Offers -10415.27089 -6119.70575
--- SqFt 41.64034 64.34714
Signif. codes: 0 *** 0.001 ** 0.01 * 0.05 . 0.1 1 Bedrooms 1083.04162 7410.54616
Bathrooms 3691.69572 12074.86126
Residual standard error: 10020 on 120 degrees of freedom
Multiple R-squared: 0.8686, Adjusted R-squared: 0.861
F-statistic: 113.3 on 7 and 120 DF, p-value: < 2.2e-16

To check if there is a premium for brick houses given everything else being equal we
test the hypothesis that β1 = 0 at the 95% confidence level. Using the regression
output we see that the 95% confidence interval for β1 is [13373.89, 21220.91]. Since
this does not include zero we conclude that brick is a significant factor when pricing
a house. Further, since the entire confidence interval is greater than zero we conclude
that people pay a premium for a brick house.
(2) To check that there is a premium for houses in Neighborhood three, given everything
else we repeat the procedure from part (1), this time looking at β3 . The regression
output tells us that the confidence interval for β3 is [14446.33, 26915.75]. Since the
entire confidence interval is greater than zero we conclude that people pay a premium
to live in neighborhood three.

14
(4) We want to determine if Neighborhood 2 plays a significant role in the pricing of
a house. If it does not, then it will be reasonable to combine neighborhoods one
and two into one “old” neighborhood. To check if Neighborhood 2 is important, we
perform a hypothesis test on β2 = 0. The null hypothesis β2 = 0 corresponds to
the dummy variable N2 being unimportant. Looking at the confidence interval from
the regression output we see that the 95% confidence interval for β2 is [−6306, 3184],
which includes zero. Thus we can conclude that it is reasonable to let β2 be zero and
that neighborhood 2 may be combined with neighborhood 1.

(3) To check that there is a premium for brick houses in neighborhood three we need to
alter our model slightly. In particular, we need to add an interaction term Brick×N 3.
This more complicated model is

Y = β0 + β1 Brick + β2 N2 + β3 N3 + β4 Bids
+ β5 SqF t + β6 Bed + β7 Bath + β8 Brick · N3 + ,  ∼ N (0, σ 2 ).

To see what this interaction term does, observe that

∂E[Y |Brick, N3 ]
= β3 + β8 Brick.
∂N3
Thus if β8 is non-zero we can conclude that consumers pay a premium to buy a brick
house when shopping in neighborhood three. The output of the regression which
includes the interaction term is below.
0.5 % 99.5 %
Coefficients: (Intercept) -19781.05615 25801.04303
Estimate Std. Error t value Pr(>|t|) BrickYes 7529.25747 20123.67244
(Intercept) 3009.993 8706.264 0.346 0.73016 N2 -6894.11333 5548.05681
BrickYes 13826.465 2405.556 5.748 7.11e-08 *** N3 8363.62557 26119.20030
N2 -673.028 2376.477 -0.283 0.77751 Offers -11187.37034 -5614.80551
N3 17241.413 3391.347 5.084 1.39e-06 *** SqFt 39.31099 68.81858
Offers -8401.088 1064.370 -7.893 1.62e-12 *** Bedrooms 588.32720 8847.99967
SqFt 54.065 5.636 9.593 < 2e-16 *** Bathrooms 823.98555 12102.74436
Bedrooms 4718.163 1577.613 2.991 0.00338 ** BrickYes:N3 -722.17781 21085.33248
Bathrooms 6463.365 2154.264 3.000 0.00329 ** 0.5 % 99.5 %
BrickYes:N3 10181.577 4165.274 2.444 0.01598 * (Intercept) -19781.05615 25801.04303
--- BrickYes 7529.25747 20123.67244
Signif. codes: 0 *** 0.001 ** 0.01 * 0.05 . 0.1 1 N2 -6894.11333 5548.05681
N3 8363.62557 26119.20030
Residual standard error: 9817 on 119 degrees of freedom Offers -11187.37034 -5614.80551
Multiple R-squared: 0.8749, Adjusted R-squared: 0.8665 SqFt 39.31099 68.81858
F-statistic: 104 on 8 and 119 DF, p-value: < 2.2e-16 Bedrooms 588.32720 8847.99967
Bathrooms 823.98555 12102.74436
BrickYes:N3 -722.17781 21085.33248

To see if there is a premium for brick houses in neighborhood three we check that
the 95% confidence interval is greater than zero. Indeed, we calculate that the 95%
confidence interval is [1933, 18429]. Hence we conclude that there is a premium at the
95% confidence level. Notice however, that the confidence interval at the 99% includes
zero. Thus if one was very stringent about drawing conclusions from statistical data,
they may accept the claim that there is no premium for brick houses in neighborhood
three.

15
Problem 8: What causes what??
Listen to this podcast:
http://www.npr.org/blogs/money/2013/04/23/178635250/episode-453-what-causes-what

1. Why can’t I just get data from a few different cities and run the regression of “Crime”
on “Police” to understand how more cops in the streets affect crime? (“Crime” refers
to some measure of crime rate and “Police” measures the number of cops in a city)

2. How were the researchers from UPENN able to isolate this effect? Briefly describe
their approach and discuss their result in the “Table 2” below.

3. Why did they have to control for METRO ridership? What was that trying to capture?
The problem here is that data on police and crime cannot tell the difference between
more police leading to crime or more crime leading to more police... in fact I would
expect to see a potential positive correlation between police and crime if looking across
different cities as mayors probably react to increases in crime by hiring more cops.
Again, it would be nice to run an experiment and randomly place cops in the streets
of a city in different days and see what happens to crime. Obviously we can’t do that!

What the researchers at UPENN did was to find a natural experiment. They were
able to collect data on crime in DC and also relate that to days in which there was a
higher alert for potential terrorist attacks. Why is this a natural experiment? Well,
by law the DC mayor has to put more cops in the streets during the days in which
there is a high alert. That decision has nothing to do with crime so it works essentially
as a experiment. From table 1 we see that controlling for ridership in the METRO,
days with a high alert (this was a dummy variable) have lower crime as the coefficient
is negative for sure. Why do we need to control for ridership in the subway? Well,
if people were not out and about during the high alert days there would be fewer
opportunities for crime and hence less crime (not due to more police). The results
from the table tells us that holding ridership fix more police has a negative impact on
crime.

Still we can’t for sure prove that more cops leads to less crime. Why? Well, imagine
the criminals are afraid of terrorists and decide not to go out to “work” during a high
alert day... this would lead to a reduction in crime that is not related to more cops in
the streets. But again, I dont believe that is a good line of reasoning so these results
are building a very strong circumstancial case that more cops reduce crime.

4. In the next page, I am showing you “Table 4” from the research paper. Just focus
on the first column of the table. Can you describe the model being estimated here?
What is the conclusion?
In table 4 they just refined the analysis a little further to check whether or not the
effect of high alert days on crime was the same in all areas of town. Using interactions
between location and high alert days they found that the effect is only clear in district
1. Again, this makes a lot of sense as most of the potential terrorists targets in DC are
in District 1 and that’s where more cops are most likely deployed to. The effect in the

16
 other districts is still negative but small and given the standard error in parenthesis
we conclude it can still be zero (why? check the confidence interval!).

effect of police on crime 271

TABLE 2
Total Daily Crime Decreases on High-Alert Days

(1) (2)
High Alert !7.316* !6.046*
(2.877) (2.537)
Log(midday ridership) 17.341**
(5.309)
R2 .14 .17

Note.—The dependent variable is the daily total number of crimes

(aggregated over type of crime and district where the crime was
Figurecommitted)
1: The independent
Washington,variable is the
D.C., during theperiod
daily total
March 12,number
2002– of crimes in D.C. This table
present the
July 30,estimated coefficients
2003. Both regressions and
contain their standard
day-of-the-week fixederrors
effects. in parenthesis. The first column
refers The
to anumber
modelofwhere
observations is 506.
the only Robust used
variable standard
in errors are inAlert dummy whereas the model
the High
parentheses.
in column (2) controls
* Significantly formfrom
different the zero
METROat the 5ridership. * refers to a significant coeficient at the
percent level.
5% level,** **
Significantly
at the 1% different
level.from zero at the 1 percent level.
274 the journal of law and economics
local officials. In addition to increasing its physical presence, the police
TABLE 4
department increases its virtual street presence by activating a closed-circuit
camera system that covers sensitive areas of the NationalonMall.
Reduction in Crime on High-Alert Days: Concentration the National
The cameraMall

system is not permanent; it is activated only during heightened terror alert

Coefficient
10
periods or during major events suchCoefficient
as presidential inaugurations.
Coefficient (Clustered by Alert
(Robust) (HAC) Status and Week)
High Alert # District 1 IV. Results
!2.621**
!2.621* !2.621*
(.044) (1.19) (1.225)
High Alert # Other Districts !.571 !.571 !.571
The results from our most basic regression
(.455)
are presented
(.366)
in Table 2,(.364)
where
we regress daily D.C. crime totals against the terror alert level (1 p high,
Log(midday ridership) 2.477* 2.477** 2.477**
0 p elevated) and a day-of-the-week
Constant
(.364)
indicator. The
!11.058**
(.522) (.527)
coefficient on!11.058
!11.058
the alert
"

level is statistically significant at the(4.211)

5 percent level
(5.87)and indicates (5.923)
that on
high-alert days, totaldependent
Note.—The crimesvariable
decrease
is daily bycrimean average
totals ofStandard
by district. sevenerrors
crimes per day,are
(in parentheses)
clustered by 6.6
or approximately district. All regressions
percent. We contain
use day-of-the-week
dummy fixed effects
variables (notandshown)
district fixedfor
effects.
each The
Figure
number of2:observations
The dependent
is 3,542. R p
2 variable
.28. HAC is p the daily total
heteroskedastic number
autocorrelation of crimes
consistent. in D.C. District 1
day of the week
refers
"
to
Significantly control
different for
from day
zero at effects
the 10 (crime
percent level. is highest on
to a dummy variable associated with crime incidents in the first police district area. Fridays).
* Significantly different from zero at the 5 percent level.
We hypothesize
This** table that
present
Significantly the level
thefrom
different zero of
estimated at thecrime
1 percentdecreases
coefficientslevel. and theiron standard
high-alert daysin in
errors parenthesis.* refers
D.C. because of greater
to a significant police at
coeficient presence
the 5% level,on the ** streets.
at the 1% Anlevel.
alternative hy-
17
pothesis isanthat tourism
official is reduced
news release on high-alert
from February 27, 2003.days, and as a result,
Unofficially, we were there
told
that during heightened alert periods, the
are fewer potential victims, which leads to fewer crimes. We are skeptical police department
11 switches from
threeexplanation
Problem
of the latter 8-hour 9:shifts
Don’t a day
on Taketo two
theoreticalYour 12-hour
Vitamins
grounds shifts,because,
thus increasing
holding the all
effective
else
police presence by 50 percent.18 Despite several requests, however, the D.C.
equal, daily
Read crimefollowing
is unlikely to vary significantly on the basis of the number
policethe would neither article:
confirm nor deny this exact procedure. Nevertheless, if
of daily visitors.
take The vast majority of visitorsfigure,
wehttp://fivethirtyeight.com/features/dont-take-your-vitamins/
50 percent as an approximate to Washington,
then we estimate D.C.,an areelasticity
never
10
of crime with respect to police presence of !15 percent/50 percent p !.3.
The increased patrols
As it turns andthis
out, activation of the
is exactly theclosed-circuit television
figure estimated system are
by Thomas discussed
Marvell and
in an official news release (see Metropolitan Police Department, MPDC Lowers Emergency
Carlisle Moody and is also consistent with a range of elasticities on different
Response Level—UPDATE (February 27, 2003) (http://mpdc.dc.gov/news/news.shtm)). We
crimes
discuss changes from approximately
in police presence in more .2 detail
to .9 analyzed by
in the text17 Levitt,
further Corman and Mocan,
below.
19
11 and Di Tella and Schargrodsky.
The premise of the argument is dubious. We spoke with people at the Washington, D.C.,
Convention and Crime mayCorporation
Tourism come in waves;
(which we control
monitors for occupancy
hotel some of this using
rates), withday-of-the-
people in
week effects,
the hotel industry, and withbutthethere
D.C. may
policebeand
other
the sources of for
statistician dependence that result
the D.C. Metro system,in
 List a few ideas/concepts that you have learned so far in this class that helps you
understand this article.

18