Title: 1. ?????? (Discrete Choice Models)
1?????????????
??? ???????????
2?????????????
- 1. ?????? (Discrete Choice Models)
- 2. ?????? (Sample Selection Models)
- 3. ?????? (Duration Models)
31 Discrete Choice Models Â
4Â
5Â
6???????
??????(1?? 0???) ??(????) 1?? 0?? ???? ????? ??? ??? ?? ?? ???
0 0 35 ? 45?
0 0 25 ? 70?
1 0 84?12? 500? 41 ? 40?
1 0 84?12? 300? 46 ? 80?
1 1 84?12? 86?6? 350? 55 ? 150?
1 1 84?12? 87?2? 1000? 49 ? 270?
7Econometric model
- Single equation model
- System of equations model
- Simultaneous equation Model
8Single equation model
Can be written as
9System of equations model
10Simultaneous equations
11Microeconometrics
- Discrete choice models
- Sample selection models
- Duration models
12 Discrete Choice Model
- Probit Model Logit Model
- Multinomial Choice Model
- Multinomial Logit Model
- Nested Logit Model
- Mixed Logit Model
- Multinomial Probit Model
- Bivariated Probit Model
- Multivariate Probit Model
- Sequential Choice Model
- Ordered Probit Model
- Count Data
13Sample Selection Model
- Censored model
- Sample selection model
14Duration Model
- Duration model
- Split population model
15Binary Choice Model
Individual i
Choose A
Dont Choose A
16(No Transcript)
17Binary Choice Model
(Unobserved variable)
18Probit Model
N(0, 1)
Assume
19Binary Choice Model
- Boczar (1978, J. of Finance)
Personal loan debtor
Bank
Finance Company
20Binary choice model
Obtain a credit from a bank
Obtain a credit from a financial company
21The probability of choosing alternative 1 is
given by
22 Probit Model
23The probability of choosing alternative 0 is
given by
24Probit Model
25Probit model
The loglikelihood for this model is given by
26Properties of Maximum Likilihood Estimator
27(No Transcript)
28Probit, logit vs. OLS
29Modeling Decision
- This yes or no type decision leads to a dummy
variable. - The dependent variable of our model is a dummy
variable. - We will be modeling the probability function,
P(Y1).
30Simplest ModelLinear Probability Model
31Picture of LPM
1
X
0
X0
X1
32Problems of LPM
- Predictions outside 0-1 range.
- Heteroscedasticity
- This can be solved and a estimated GLS estimator
developed. - Coefficient Determination has little meaning.
- Constant marginal effect.
33Interpreting the Probit Model
34The logit model
35The Log-Likelihood function
36LIMDEP Command  Read NVAR7Nobs200
filenames.. Â Regress LHSy1
RHSone,x1,x2,. Probit LHSy1
RHSone,x1,x2,. Logit LHSy1
RHSone,x1,x2,.
37PROBIT, LOGIT Goodness of Fit Measures?
- More often cited are R-square values based on
likelihood ratios. - Maddala Â
- R2 1 - (LR / LUR) 2/n
- McFadden R-square
- R2 1 - (log(LUR ) / log(LR))
38Jacobson and Roszbach (2003, Journal of Banking
Finance) ----- Bivariate Probit Model
Providing a loan?
Loan defaults?
Yes
No
Yes
No
39Bivariate Probit Model
(if loan granted)
(if loan not granted)
(if loan does not default)
(if loan defaults)
40Bivariate Probit model
41Multivariate Probit Model
Cigarette
Alcohol
Marijuana
Cocaine
YES
No
No
YES
YES
YES
No
No
42(No Transcript)
43Multivariate Normality
44Hausman and Wise (1978, Econometrica)
45Multivariate Probit Model
- J3, Clark (1961)
- J4, Hausman and Wise (1978, Econometrica)
- J gt 4
- McFadden (1989, Econometrica)
- ------ Simulation-Based Estimation
- ------ high dimensional integrals
- Stern (1997, Journal of Economic Literature)
- ----- Simulated Maximum Likelihood Estimator
- ----- Simulated Moment Estimator
- ----- GHK simulator
46LIMDEP CommandBivariated Probit Model  Read
NVAR7Nobs200 filenames.. Â Bivariate
Probit LHSY1, Y2
RHSone,x1,x2,.
RH2one,z1,z2,.
47Multivariate Probit Model
48Multinomial Choice Model Example Credit Card
Individual i
Alternatives
J
2
3
1
49Multinomial Choice Model
50Multinomial Choice Model
51Multinomial Logit Model
Let
be the probabilities associated these m categories
( j1,2,.m-1)
52If
53McFadden 1973
54Multinomial Logit Model
55If the ith individual falls in the jth category
otherwise
56Independence of Irrelevant Alternatives (IIA)
57Ordered Probit Model
Example Blume, Lim, Mackinlay (1998, Jornal of
Finance) Corporate bond rating (????)
AAA
AA
A
BBB
58Ordered Probit Model
N(0, 1)
59(No Transcript)
60????
?
????
?
????
?????? Never Fail
????? Eventually Fail
61Sequential Choice Model Example???
Auction
No
Yes
No
Yes
Yes
No
62Sequential Response Model
63Sequential Choice Model
First Auction
No 1-F(ß1x)
YesF(ß1x)
No1-F(ß2x)
YesF(ß2x)
YesF(ß3x)
No1-F(ß3x)
64Then the probabilities can be written as
65Model Selection Joint decision vs.
Sequential decision
EXAMPLE
Bivariate Probit Model ? Multinomial Choice
Model? Ordered Probit Model? Sequential
Choice Model?
66Model Selection
EXAMPLE
Ioannides and Rosenthal (1994, The Review of
Economics and Statistics) Estimating the
consumption and investment demand for housing and
their effect on housing tenure status
67Multinomial Choice Model?
(?????)
(??????)
(??????)
(???????)
68Ordered Probit Model?
Intensity of Utility
(???????)
(??????)
(??????)
(?????)
69Sequential Choice Model?
???
??
??
???
70Bivariate Probit Model ?
??
???
??
???
71Bivariate Probit Model
72Bivariate Probit model
73Multivariate Probit Model
Cigarette
Alcohol
Marijuana
Cocaine
YES
No
No
YES
YES
YES
No
No
74(No Transcript)
75Multivariate Normality
76Multivariate Probit Model
- J3, Clark (1961)
- J4, Hausman and Wise (1978, Econometrica)
- J gt 4
- McFadden (1989, Econometrica)
- ------ Simulation-Based Estimation
- ------ high dimensional integrals
- Stern (1997, Journal of Economic Literature)
- ----- Simulated Maximum Likelihood Estimator
- ----- Simulated Moment Estimator
- ----- GHK simulator
77ApplicationSurvey bias models
- Censored model Deaton and Irish (1984)
- Probit model Bollinger and David (1997),
Abrevaya and Hausman (1999a) - Multinomial logit model Hsiao and Sun (1999)
- Ordered probit model Dustman and van Soest
(2004) - Duration model Torelli and Trivellato (1993),
Abrevaya and Hausman (1999b)
78Bollinger and David (1997)
Did not use the substance
Used the substance
Lied
Did not lie
Lied
Did not lie
79Hsaio and Sun (1999)
Did not use the substance
Used the substance
lied
Did not lie
80 Table 1 Replied
Substance Use qi qi yi(1 - wi
)
yi 1 yi 0
(used the substance) (did not use the substance)
wi 1 (lied) 0 Not Applicable
wi 1 (lied) (reply did not use) Not Applicable
wi 0 (did not lie) 1 0
wi 0 (did not lie) (reply used) (reply did not use)
81Model 1 Uniform one-sided survey response bias
model
The loglikelihood for this model is given by
82Model 2 Heterogeneous and independent one-sided
survey response bias model
83Model 3 Heterogeneous and dependent one-sided
survey response bias model
Assume
84Partial observability model Poirier (1980)
85Partial observability model Poirier (1980)
- Zero-inflated Poisson model
- Double hurdle model
- Split population model
86Leung and Yu(2003) Empirical Analysis
- 1998 U.S. National Household Survey on Drug Abuse
(NHSDA) - Tobacco
- Alcohol
- Marijuana
- Cocaine
87Table 4 Model 1 Uniform One-sided Survey
Response Bias Model
Variable Tobacco Alcohol Marijuana Cocaine
------- ---- ---- ----
------- ---- ---- ----
a 0.1386 0.1614 0.0003 0.4701
-0.0586 -0.0192 -0.3202 -0.7591
Loglikelihood -13852.738 -15181.128 -6073.143 -1519.109
88Count regression
- Appropriate when the dependent variable
- is a non-negative integer (0,1,2,3,)
89- Distributions and Models
- Poisson Model
- Negative Binomial Model
- Zero-inflated Poisson Model
- Zero-inflated Negative Binomial Model
90Poisson Regression
91Why not use linear regression?
- Typical count data in health care
- Large number of 0 values and small values
- Discrete nature of data
- Result
- Unusual distribution
92Normal distribution vsPoisson distribution
Bell shaped curve
Normal distribution
Poisson distribution
Not bell shapednext slide
Intensity of process
93Poisson with ? 0.5
94When Count Data Cannot be Treated Normally
95When they probably can.
96What happens when mean ? variance?
- Overdispersion when variance gt mean
- Sometimes called unobserved heterogeneity
- Zero-Inflated More zeros than expected by
Poisson distribution - Ex. If ?1 (mean1), then we expect 37 0s
97Overdispersion
98Poisson Regression models
Negative Binomial Regression models
u is Weibull distribution
99Overdispersion and Zero Inflation
100Zero-inflated Poisson
101Example
- Bao article
- Predicting the use of outpatient mental health
services do modeling approaches make a
difference? Inquiry. 2002 Summer39(2)168-83.
102Observed data
103Poisson and Zero-Inflated Poisson
104Negative Binomial Model
105Zero-Inflated Negative Binomial Model
106TOBIT Model
107TOBIT MODEL
if
108TOBIT MODEL
109TOBIT MODEL
110TOBIT MODEL
111TOBIT MODEL
112TOBIT MODEL
113TOBIT MODEL
114TOBIT MODEL
where
p.f.
115TOBIT MODEL
let
116TOBIT MODEL
NOTE
117TOBIT MODEL
let
118TOBIT MODEL
119TOBIT MODEL
120The log-likelihood function
121Sample Selection Model
122Self- Selection Model
123Sample Selection Model
124Heckmans Two-step Estimator (1979)
125Application Heckmans Two-step Estimator vs.
Two-part model
Duan, Manning, Morris, Newhouse(1984,
JBES) Maddala (1985a, 1985b) Hay, Leu,
Fohrer(1987,JBES) Manning, Duan, Rogers (1987,
JE) Leung and Yu (1996, JE) Dow and Norton (2003,
HSORM) Dow and Norton (2005)
126Sequential Choice Model?
???
??
??
???
0
127Joint Choice Model
???
??
???
??
128Duration Models
- ? Censored Data
- ? Unobserved Heterogeneity
- ? Time-Varying Covariates
129 D
C
C
D
D
End of study
130 131 Hazard Rate and Survival Rate
132Duration Model
133- Distributions
- Parametric
- Expoential
- Weibull
- Log-normal
- Log-logistic
- Gamma
- Semi-parametric
- Coxs partial likelihood estimator
134LIMDEP Command---Duration Model  Read
NVAR7Nobs200 filenames.. Â Survival
LHSln(time), status (exit1) RHSone,x1,x2,.
modelExponential Survival LHSln(time),
status (exit1) RHSone,x1,x2,.
modelWeibull Coxs Semiparametric
Estimator Survival LHSln(time), status
(exit1) RHSone,x1,x2,.
135TOBIT MODEL
if
136TOBIT MODEL
137TOBIT MODEL
138TOBIT MODEL
139TOBIT MODEL
where
p.f.
140TOBIT MODEL
let
141TOBIT MODEL
NOTE
142TOBIT MODEL
let
143TOBIT MODEL
144TOBIT MODEL
NOTE
by LHopital rule
145Duration Model
D
C
C
D
D
End of study
146Duration model
- Censored data
- Unobserved heterogeneity
- Time-varying covariates
1472.2 Hazard Analysis
148Survival rate and Hazard rate
1492.2 Nonparametric Hazard Analysis
- Kaplan-Meier estimator
- Life table estimator
150Figure 2 Kaplan-Meier Estimates of Survival
Function
151Figure 3 Life Table Estimates of Survival
Function
152The density and survival functions
f (Ti?wi) the probability density function
of the failure time S
(ti?wi) the probability of survival
153The specifications for f (Ti?xi) and S
(ti?xi)
- Exponential
- Weibull
- Log-logistic
- Log-normal
154Figure 4 Life Table Estimates of Hazard
Functions
155Eventually fail assumption
156????
?
?
????
????
Eventually Fail Assumption
157????
?
????
?
????
?????? Never Fail
????? Eventually Fail
158Schmidt and Witte (1989) --- Split
population duration model
G (?xi) the probability of eventual failure f
(Ti?wi) the probability density function
of the failure time S (Ti?wi) the
probability of survival
159Schmidt and Witte (1989) --- Likelihood function
1604.3 Multivariate Split Population Duration Model
161Multivariate probit model
162Multivariate duration model
163Unobserved heterogeneity
The frailty ( m 1,2) is assumed to follow
a gamma distribution with mean 1 and variance
164Whether Part
- individuals probability
of eventual failure for a type k event (k
1,2,3,4). - follows a Weibull distribution
165Duration Part
- Assume the survival function is log-logistic. The
second frailty - enters the hazard function as
, and is the failure time or the
where
censored time, whichever is earlier.
166The cumulative hazard, the survival function, and
the density function are
167The likelihood function is given by
168Partial likelihood Estimation
169Person Event Time Li
1 1 3
2 2 8
3 10
4 15
5 3 21
6 4 30
7 5 32
8 52
9 52
10 52
170B.3 Simultaneous Equations Models
- M. J. Lee (1995, Journal of Applied Econometrics)
171