Scribd
Upload
47 views26 pages

Econometric Lec7

This document discusses qualitative response regression models, where the dependent variable is categorical with two or more categories. It covers linear probability models, logit models, probit models, and multinomial logit models. Key points include: the linear probability model relates probability to independent variables linearly; the logit and probit models use cumulative distribution functions to relate probability to the independent variables nonlinearly; maximum likelihood is used to estimate model parameters for logit/probit models. Examples are provided to illustrate estimation and interpretation.

Uploaded by

nhung
Copyright
© © All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as PPT, PDF, TXT or read online on Scribd
47 views26 pages

Econometric Lec7

This document discusses qualitative response regression models, where the dependent variable is categorical with two or more categories. It covers linear probability models, logit models, probit models, and multinomial logit models. Key points include: the linear probability model relates probability to independent variables linearly; the logit and probit models use cumulative distribution functions to relate probability to the independent variables nonlinearly; maximum likelihood is used to estimate model parameters for logit/probit models. Examples are provided to illustrate estimation and interpretation.

Uploaded by

nhung
Copyright
© © All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as PPT, PDF, TXT or read online on Scribd
You are on page 1/ 26

TOPICS IN

ECONOMETRICS

7
LECTURE

1
CHAPTER 15

Qualitative Response
Regression Models

2
The Nature of Qualitative
Response Models

The dependent variable is categorical

- Binary or Dichotomous: 2 categories

- Polychotomous or multiple-category:
more than 2 categories

3
Examples

- Smoking or not smoking

- Owning a house or not Owning a house

- Participating the labor force or not

- Selecting the brands of products A, B or C

- Voting for Parties A, B or C


4
Methods of Modeling
• We will consider the dependence of Probability of
something happening on independence variables.

• Some Models can be used:

- The Linear Probability Models (LPM)


- The Logit Model (Logistic Regression)
- The Probit Model (CDF Normal)
- The Multinomial Model
5
The Linear Probability Model (LPM)

• Assume that the dependent variable has two


categories:
- A family owns a house
- A family does not own a house
Let Yi=1 if owns a house and Yi=0 otherwise
Yi follows Bernoulli Distribution with Probability Pi

P(Yi=1)= Pi
E(Yi/Xi)= Pi 6
The Linear Probability Model (LPM)

• Modeling:The Linear Probability Models


E(Yi/Xi)= Pi= β1 + β2Xi

Condition: 0 ≤ E(Yi/Xi) ≤ 1

• Some Problems:
- Non-Normality of ui
- Heteroscedasticity
- Nonfulfillment of 0 ≤ E(Yi/Xi) ≤ 1 7
The Linear Probability Model (LPM)

- Value of R2

8
The Linear Probability Model (LPM)

Example

Table 15.1: Estimate LPM for the data of the


families owning a house or not depending on
Income

9
The Logit Model

• Normally, Pi is nonlinear related to Xi

• We better use the Cumulative Distribution Function10


The Logit Model
• Logistic Distribution Function

exp( 1   2 X i )
Pi 
1  exp( 1   2 X i )
• Odds ratio:
Pi
 e 1   2 X i
1  Pi

• Log-odds Ratio form:


Pi
Li  ln  1   2 X i
1  Pi 11
The Logit Model - Interpretation

• Slope coefficient measures how L change for 1


unit change in X. L means the log of the odds
in favor of Y=1. Then, the odds of Y=1
changes by exp(β2)

• dP/dX=β2 P(1-P): then, if we want to know how


P changes when X change, we need to know
P change at what level.
12
The Logit Model - Estimation

• We can not use OLS method to estimate


parameters as if Y=1, we have L=ln(1/0) and
Y=0, L=ln(0/1). For logistic regression, we
have to use Maximum Likelihood (ML) method
to estimate the parameters.
(see appendix 15A for more details)

• The value of log likelihood function is similar to


RSS in OLS.
13
The Logit Model - Inference

• As we use (ML) method, the sample should be large.


Then, instead of using t-stats to test for significance
of parameters, we use z-test.

• As for binary data, R2 defined as ESS/TSS is


normally not meaningful. We will use pseudo R2 to
measure the goodness of fit of the model.
- McFadden R2 defined as 1-[lnL(Mfull)/lnL(Mintercept))]
- Count R2 defined as #of correct predictors/n
in which predicted Prob > 0.5 will be classified as 1
predicted Prob < 0.5 will be classified as 0
14
The Logit Model - Inference

• To test for the significance of the model (similar


to F test in OLS method), we use Likelihood
Ratio (LR) statistic. LR follows chi square
Distribution with df=#of independent variables

• LR= -2ln[L(Mfull)/L(Mintercept))]

15
The Logit Model - Examples

• Example for the table 15.7

• Problem 15.7 book page 583

• Problem 15.9 book page 583

16
The Probit Model
• We use CDF normal to construct the model for
Probability Pi

• Pi=P(Y=1)=F(β1 + β2Xi)

• Let Ii=β1 + β2Xi,, Ii are called utility index

• Ii= F-1(Pi)= β1 + β2Xi

• dP/dX=f(β1 + β2Xi) β2 . This measures how P change for


1 unit change of X in which f(.) is the normal density
function
17
The Probit Model

18
The Probit Model

CDF normal

19
The Probit Model

Examples

- Redo the example for the table 15.7 using


probit model.

- Problem 15.19 on page 589.

20
Multinomial Logit Model
• Assume we have J choices A0, A1, A2...Aj,
J>2 (J categories).
• We want to modeling the probability
choice Aj will be selected.
• Extend the logit model for Binary data
• Select one category as the baseline.
• The logistic model of the category j is:

21
Multinomial Logit Model
• For category j
exp(1 j   2 j X ij )
Pij 
1  exp(1 j   2 X ij )
For baseline category

1
Pij 
1  exp(1 j   2 X ij )
22
Log-odds Ratio Form

Pij
L j  ln  1 j   2 j X ij
1  Pij

23
Model for Count Data
Example:
- Number of registered students to a course per year
- The number of visits to the hospital per year
- The number of cars parking with 1 hour

We want to study the dependence of these above


factors on some other factors  Use the regression
model for count data

24
Model for Count Data
Poisson Distribution:
- Probability mass function
P(Y=yi)= µyiexp(-µ)/(yi)!

E(Y)= µ and V(Y)= µ

- Linear regression model:

E(Yi)= µi = β1+ β2X2i+...+ βkXki


25
THE END

26

You might also like