Regression Models

Part 9: Model Building9-1/43

Regression ModelsProfessor William GreeneStern School of Business

IOMS DepartmentDepartment of Economics

Part 9: Model Building9-2/43

Regression and Forecasting Models

Part 9 – Model Building

Part 9: Model Building9-3/43

Multiple Regression Models

Using Binary Variables Logs and Elasticities Trends in Time Series Data Using Quadratic Terms to Improve the Model

Part 9: Model Building9-4/43

Using Dummy Variables Dummy variable = binary variable

= a variable that takes values 0 and 1. E.g. OECD Life Expectancies compared to the

rest of the world:DALE = β0 + β1 EDUC + β2 PCHexp + β3 OECD + ε

Australia, Austria, Belgium, Canada, Czech Republic, Denmark, Finland, France, Germany, Greece, Hungary, Iceland, Ireland, Italy, Japan, Korea, Luxembourg, Mexico, The Netherlands, New Zealand, Norway, Poland, Portugal, Slovak Republic, Spain, Sweden, Switzerland, Turkey, United Kingdom, United States.

Part 9: Model Building9-5/43

OECD Life Expectancy

According to these results, after accounting for education and health expenditure differences, people in the OECD countries have a life expectancy that is 1.191 years shorter than people in other countries.

Part 9: Model Building9-6/43

A Binary Variable in Regression

We set PCHExp to 1000, approximately the sample mean.

The regression shifts down by 1.191 years for the OECD countries

Part 9: Model Building9-7/43

Dummy Variable in a Log Regression

E.g., Monet’s signature equation

Log$Price = β0 + β1 logArea + β2 Signed

Unsigned: PriceU = exp(α) Areaβ1

Signed: PriceS = exp(α) Areaβ1 exp(β2)Signed/Unsigned = exp(β2)%Difference = 100%(Signed-Unsigned)/Unsigned

= 100%[exp(β2) – 1]

Part 9: Model Building9-8/43

The Signature Effect: 253%

100%[exp(1.2618) – 1] = 100%[3.532 – 1] = 253.2 %

Part 9: Model Building9-9/43

Monet Paintings in Millions

Square Inches

Price

70006000500040003000200010000

30

25

20

15

10

5

0

01

Signed

Scatterplot of Price vs Square Inches

Predicted Price is exp(4.122+1.3458*logArea+1.2618*Signed) / 1000000

Difference is about 253%

Part 9: Model Building9-10/43

Logs in Regression

Part 9: Model Building9-11/43

Elasticity

The coefficient on log(Area) is 1.346 For each 1% increase in area, price goes up by

1.346% - even accounting for the signature effect. The elasticity is +1.346 Remarkable. Not only does price increase with

area, it increases much faster than area.

Part 9: Model Building9-12/43

Monet: By the Square Inch

Area

price

70006000500040003000200010000

20000000

15000000

10000000

5000000

0

Scatterplot of Price vs Area

Part 9: Model Building9-13/43

Logs and Elasticities

Theory: When the variables are in logs:

change in logx = %change in x

log y = α + β1 log x1 + β2 log x2 + … βK log xK + ε

Elasticity = βk

Part 9: Model Building9-14/43

Elasticities

Price elasticity = -0.02070 Income elasticity = +1.10318

Part 9: Model Building9-15/43

A Set of Dummy Variables

Complete set of dummy variables divides the sample into groups.

Fit the regression with “group” effects. Need to drop one (any one) of the

variables to compute the regression. (Avoid the “dummy variable trap.”)

Part 9: Model Building9-16/43

Rankings of 132 U.S.Liberal Arts CollegesReputation = β0 + β1Religious + β2GenderEcon + β3EconFac + β4North + β5South + β6Midwest + β7West + ε

Nancy Burnett: Journal of Economic Education, 1998

Part 9: Model Building9-17/43

Minitab does not like this model.

Part 9: Model Building9-18/43

Too many dummy variables If we use all four region dummies, a is reduntant

Reputation = b0 + bn + … if north Reputation = b0 + bm + … if midwest Reputation = b0 + bs + … if south Reputation = b0 + bw + … if west

Only three are needed – so Minitab dropped west Reputation = b0 + bn + … if north Reputation = b0 + bm + … if midwest Reputation = b0 + bs + … if south Reputation = b0 + … if west

Part 9: Model Building9-19/43

Unordered Categorical Variables

House price data (fictitious)Style 1 = Split levelStyle 2 = RanchStyle 3 = ColonialStyle 4 = TudorUse 3 dummy variables for this kind of data. (Not all 4)Using variable STYLE in the model makes no sense. You could change the numbering scale any way you like. 1,2,3,4 are just labels.

Part 9: Model Building9-20/43

Transform Style to Types

Part 9: Model Building9-21/43

Part 9: Model Building9-22/43

House Price Regression

Each of these is relative to a Split Level, since that is the omitted category. E.g., the price of a Ranch house is $74,369 less than a Split Level of the same size with the same number of bedrooms.

Part 9: Model Building9-23/43

Better Specified House Price Model

Part 9: Model Building9-24/43

Time Trends in Regression

y = β0 + β1x + β2t + ε β2 is the year to year increase not explained by anything else.

log y = β0 + β1log x + β2t + ε (not log t, just t) 100β2 is the year to year % increase not explained by anything else.

Part 9: Model Building9-25/43

Time Trend in Multiple Regression

After accounting for Income, the price and the price of new cars, per capita gasoline consumption falls by 1.25% per year. I.e., if income and the prices were unchanged, consumption would fall by 1.25%. Probably the effect of improved fuel efficiency

Part 9: Model Building9-26/43

A Quadratic Income vs. Age Regression+----------------------------------------------------+| LHS=HHNINC Mean = .3520836 || Standard deviation = .1769083 || Model size Parameters = 3 || Degrees of freedom = 27323 || Residuals Sum of squares = 794.9667 || Standard error of e = .1705730 || Fit R-squared = .7040754E-01 |+----------------------------------------------------++--------+--------------+--+--------+|Variable| Coefficient | Mean of X|+--------+--------------+-----------+ Constant| -.39266196 AGE | .02458140 43.5256898 AGESQ | -.00027237 2022.85549 EDUC | .01994416 11.3206310+--------+--------------+-----------+

Note the coefficient on Age squared is negative. Age ranges from 25 to 65.

Part 9: Model Building9-27/43

Implied By The Model

Part 9: Model Building9-28/43

A Better Model?

Log Cost = α + β1 logOutput + β2 [logOutput]2 + ε

Part 9: Model Building9-29/43

Candidate Models for CostThe quadratic equation is the appropriate model.

Logc = a + b1 logq + b2 log2q + e

Part 9: Model Building9-30/43

27,326 Household Head Interviews in Germany, 1984 – 1994.

Part 9: Model Building9-31/43

Interaction Term

Education

Age*Education

Part 9: Model Building9-32/43

0 1 2 3

1 3

1

3

logIncome = β +β Educ+β Age+β Age×Educ+...+εEffect of a year of Educ depends on AgedlogIncome/dEduc = β +β Ageb = -0.022385b = 0.0019006Age = 21, elasticity = 0.017528Age = 35, elasticity = 0.044146

Part 9: Model Building9-33/43

Case Study Using A Regression Model: A Huge Sports Contract

Alex Rodriguez hired by the Texas Rangers for something like $25 million per year in 2000.

Costs – the salary plus and minus some fine tuning of the numbers

Benefits – more fans in the stands. How to determine if the benefits exceed the

costs? Use a regression model.

Part 9: Model Building9-34/43

PDV of the Costs

Using 8% discount factor Accounting for all costs Roughly $21M to $28M in each year from

2001 to 2010, then the deferred payments from 2010 to 2020

Total costs: About $165 Million in 2001 (Present discounted value)

Part 9: Model Building9-35/43

Benefits More fans in the seats

Gate Parking Merchandise

Increased chance at playoffs and world series Sponsorships (Loss to revenue sharing) Franchise value

Part 9: Model Building9-36/43

How Many New Fans? Projected 8 more wins per year. What is the relationship between wins

and attendance? Not known precisely Many empirical studies (The Journal of

Sports Economics) Use a regression model to find out.

Part 9: Model Building9-37/43

Baseball Data 31 teams, 17 years (fewer years for 6 teams) Winning percentage: Wins = 162 * percentage Rank Average attendance. Attendance = 81*Average Average team salary Number of all stars Manager years of experience Percent of team that is rookies Lineup changes Mean player experience Dummy variable for change in manager

Part 9: Model Building9-38/43

Baseball Data (Panel Data – 31 Teams, 17 Years)

Part 9: Model Building9-39/43

A Regression Model

0,team

1

Attendance(team,this year) =

+ γ Attendance(team, last year) + β Wins (team,this year)

2

3

+ β Wins(team, last year) + All_Stars(team, this year)

+ (team, this year)

Part 9: Model Building9-40/43

A Dynamic Equationy(this year) = f[y(last year)…]

0 1Fans(t)=b +b Wins(t)+cFans(t-1)+ (Loyalty effect)Suppose Fans(0) = Fans0 (Start observing in a base year)Suppose we fix Wins(t) at some Wins* and at 0 (no information).What values

0 1

0 1 0 1

0 1 0 1 0 1

0

does Fans(t) take in a sequence of years?Fans(1) = b + b Wins* + cFans0Fans(2) = b + b Wins* + c(b + b Wins* + cFans0)Fans(3) = b + b Wins* + c(b + b Wins* + c(b + b Wins* + cFans0))Fans(4) = b 1 0 1 0 1 0 1

2 t-1 2 t-1 t0 1

+ b Wins* + c(b + b Wins* + c(b + b Wins* + c(b + b Wins* + cFans0)))etc.

Collect terms: Fans(t) = b (1+c+c ... c ) b Wins*(1+c+c ... c )+c Fans0Suppose 0 < c < 1.

Fans finally settles down at

0 1 1b b b dFans* Fans* = + Wins*. = 1-c 1-c 1-c dWins *

Part 9: Model Building9-41/43

Marginal Value of One More Win

0 1 2 3

0 1 2 3

Our Model is Fans(t) = + β Wins(t) + β Wins(t-1) + β AllStars + γFans(t-1)Using the formula for the value of Fans*

+β Wins*+β Wins*+β AllStarsFans*=

1-γ

The effect of one more Win every year would b

1 2

3

e dFans*/dWins* = 1

The new player will definitely be an All Star, so we add this effect as well.The effect of adding an All Star player to the team would be / (1 )

Part 9: Model Building9-42/43

= .54914

1 = 11093.7

2 = 2201.2

3 = 14593.5

Effect of 1 more win11093.7 2201.2= 32757

1 .59414Effect of adding an All Star

14593.5= 359571 .59414

Part 9: Model Building9-43/43

Marginal Value of an A Rod 8 games * 32,757 fans + 1 All Star = 35957

= 298,016 new fans 298,016 new fans *

$18 per ticket $2.50 parking etc. $1.80 stuff (hats, bobble head dolls,…)

About $6.67 Million per year !!!!! It’s not close.

(Marginal cost is at least $16.5M / year)