7/9/2023

National Economics University

E-PhD, Cohort 6

E-PhD3121: Econometrics

LECTURE 6&7: QUALITATIVE AND

LIMITED DEPENDENT VARIABLE

Bach Ngoc Thang

[email protected]

Hanoi, July 2023

Tentative outlines
• Binary outcome models:
• For choice behavior.
• Probit model (transport.dta).
• Some reference to logit model.
• Tobit and selection models:
• Censored data.
• Tobit model interpretation (mroz.dta).
• Sample selection (mroz.dta).
• Practice and extensions:
• The transportation equation: private vs. public transport.
• Labor supply and wages of married women.
7/9/2023 2

1
 7/9/2023

Binary outcome models

• Designed for choice behavior exhibited by individuals, or a
decision process.
• Some examples:
• Decision to pursue higher education, especially among
female lecturers at Ph.D. level.
• Why some individuals take a second or third job.
• Why some bank loan applications are accepted, while
others are not.
• The effect of trade unions on firms’ training and
innovation activities.
• The choice behavior of recent higher-educational
graduates toward the informal sector.
7/9/2023 3

Specification of binary choice models:

The transportation problem
• An individual’s choice:
1 𝑖𝑛𝑑𝑖𝑣𝑖𝑑𝑢𝑎𝑙 𝑑𝑟𝑖𝑣𝑒𝑠 𝑡𝑜 𝑤𝑜𝑟𝑘
𝑦={
0 𝑖𝑛𝑑𝑖𝑣𝑖𝑑𝑢𝑎𝑙 𝑡𝑎𝑘𝑒𝑠 𝑏𝑢𝑠 𝑡𝑜 𝑤𝑜𝑟𝑘
• The probability function:
𝑓 𝑦 = 𝑝 𝑦 (1 − 𝑝)1−𝑦 , 𝑦 = 0,1 (1𝑎)
Where, y is a random variable; p is the probability that y takes the
value one; E(y) = p and Var(y) = p(1-p).
• The linear probability model:
𝑦 = 𝐸 𝑦 + 𝑒 = 𝑝 + 𝑒 (1)
𝐸 𝑦 = 𝑝 = 𝛽1 + 𝛽2 𝑥 (2)
𝑦 = 𝐸 𝑦 + 𝑒 = 𝛽1 + 𝛽2 𝑥 + e (3)
Where, e is the usual random error term, that is i.i.d;
𝑉𝑎𝑟 𝑦 = 𝛽1 + 𝛽2 𝑥 1 − 𝛽1 − 𝛽2 𝑥 (4)

7/9/2023 4

2
 7/9/2023

The probit model: standard normal

cumulative distribution function

7/9/2023 5

The probit model: standard normal

probability density function

7/9/2023 6

3
 7/9/2023

The probit model:

probability and distribution functions
• The probability density function:
1 −0.5𝑧 2
φ(z)= 𝑒 (5)
2𝜋
• The cumulative distribution function (c.d.f), or the probit
function:
𝑧
1 −0.5𝑢2
Φ(z)=P[Z ≤ 𝑧] = න 𝑒 𝑑𝑢 (6)
−∞ 2𝜋
• The probit statistical model:
𝑝 = 𝑃 𝑍 ≤ 𝛽1 + 𝛽2 𝑥 = Φ 𝛽1 + 𝛽2 𝑥 (7)

7/9/2023 7

Interpretation of the Probit model

• The marginal effect:
𝑑𝑝 𝑑Φ(t) 𝑑𝑡
= 𝑑𝑡 . 𝑑𝑥 = φ 𝛽1 + 𝛽2 𝑥 𝛽2 (8)
𝑑𝑥
Where, 𝑡 = 𝛽1 + 𝛽2 𝑥.
• Three implications of equation (8):
• The sign of the marginal effect is conditional on 𝛽2;
• When 𝑝 = Φ 0 = 0.5, the effect of a change in 𝑥 has its
greatest effect – the individual is “on the borderline”; and
• The other two polar cases:
• When 𝛽1 + 𝛽2 𝑥 is positively large, near 3, the
probability to drive (private car) is close to one.
• When 𝛽1 + 𝛽2 𝑥 is negatively large, near -3, the
probability to drive (private car) is close to zero.
7/9/2023 8

4
 7/9/2023

Maximum likelihood estimation of

the Probit model
• From equations (1a) & (7), the probability function for 𝑦:
𝑓 𝑦𝑖 = [Φ 𝛽1 + 𝛽2 𝑥 ]𝑦𝑖 [1 − Φ 𝛽1 + 𝛽2 𝑥 ]1−𝑦𝑖 , 𝑦𝑖
= 0,1 (9)
• The joint probability density function for y1, y2 , y3:
𝑓 𝑦1 , 𝑦2 , 𝑦3 = 𝑓(𝑦1 )𝑓(𝑦2 )𝑓(𝑦3 )
• A specific case: y1 = 1, y2 = 1, y1 = 0:
𝑃 𝑦1 = 1, 𝑦2 = 1, 𝑦3 = 0 = 𝑓 𝟏, 𝟏, 𝟎
= 𝑓 1 𝑓 1 𝑓 0 (10)
• With x1 = 15, x2 = 6, x3 = 7:
𝑃 𝑦1 = 1, 𝑦2 = 1, 𝑦3 = 0 = Φ 𝛽1 + 𝛽2 (15) . Φ(𝛽1 +
𝛽2 (6)). 1 − Φ 𝛽1 + 𝛽2 7 = 𝐿(𝛽1 , 𝛽2 ) (11)
➔ The likelihood function.
7/9/2023 9

The log-likelihood function

𝑙𝑛𝐿(𝛽1 , 𝛽2 ) = ln൛Φ 𝛽1 + 𝛽2 15 . Φ൫𝛽1 +
𝛽2 6 ൯. 1 − Φ 𝛽1 + 𝛽2 7 ൟ = 𝑙𝑛Φ൫𝛽1 +
𝛽2 15 ൯ + 𝑙𝑛Φ 𝛽1 + 𝛽2 6 + lnൣ1 − Φ൫𝛽1 +
𝛽2 7 ൯൧ (12)
• Maximizing (12) yields the maximum likelihood
estimates of 𝛽1 and 𝛽2 .
• In large samples the maximum liekelihood estimator
is normally distributed, consistent and best.

7/9/2023 10

10

5
 7/9/2023

The transportation data: transport.dta

7/9/2023 11

11

Probit estimates of the transport data

7/9/2023 12

12

6
 7/9/2023

Marginal effect at a specific value

• At Dtime= 2 (20 minutes):
෢
𝑑𝑝
= φ(𝛽෨1 + 𝛽෨2 𝐷𝑡𝑖𝑚𝑒)𝛽෨2
𝑑𝐷𝑡𝑖𝑚𝑒
= φ(−0.0644+0.3x2)(0.3)=φ(0.5355)(0.3) =0.3456x0.3=0.1037
See slide 6 for the density function evaluated at φ(0.5355).

• The probability at Dtime = 3:

𝑝Ƹ = Φ(𝛽෨1 + 𝛽෨2 𝐷𝑡𝑖𝑚𝑒) = Φ −0.0644 + 0.3𝑥3 = 0.7983

7/9/2023 13

13

The average marginal effect

7/9/2023 14

14

7
 7/9/2023

Censored data: Histogram of wife’s

hours of work in 1975

7/9/2023 15

15

OLS estimates with uncensored

sample data

7/9/2023 16

16

8
 7/9/2023

OLS estimates with censored sample

data

7/9/2023 17

17

The Tobit model: Maximum

likelihood estimator
• The full likelihood function is the product of the
probabilities that the limit observations occurs times
the probability density functions for all the positive,
non-limit, observations (see textbook).
• The marginal effect (left-censored case at 0):
𝜕𝐸 𝑦 𝑥 𝛽1 + 𝛽2 𝑥
= 𝛽2 Φ (13)
𝜕𝑥 𝜎
Where, the c.d.f. values are positive, the sign effect
is conditional on the estimated value of 𝛽2 . 𝜎 is the
standard deviation (s.d.) of the error term 𝑒𝑖 .
7/9/2023 18

18

9
 7/9/2023

The Tobit model: labor supply, work

hour data, by married women
• The baseline model:
𝐻𝑜𝑢𝑟𝑠 = 𝛽1 + 𝛽2 𝐸𝑑𝑢𝑐 + 𝛽3 𝐸𝑥𝑝𝑒𝑟 + 𝛽4 𝐴𝑔𝑒 +
𝛽5 𝐾𝑖𝑑𝑠𝑙6 + 𝑒 (14)

7/9/2023 19

19

Labor supply and wage data, by

married women (mroz.dta)

7/9/2023 20

20

10
 7/9/2023

Labor supply and wage data, by

married women: Summary statistics

7/9/2023 21

21

Tobit estimates of labor supply, by

married women

7/9/2023 22

22

11
 7/9/2023

Marginal effect at the mean

•For the mean values: Educ = 12.29, Exper =
10.63, Age= 42.54, Kidsl6 = 1 (not 0.24):
෣ = 𝛽መ1 + 𝛽መ2 𝐸𝑑𝑢𝑐 + 𝛽መ3 𝐸𝑥𝑝𝑒𝑟 + 𝛽መ4 𝐴𝑔𝑒 +
• 𝐻𝑜𝑢𝑟𝑠
𝛽መ5 𝐾𝑖𝑑𝑠𝑙6 =1349.9+73.3x12.29+80.5x10.63
− 60.8𝑥42.54 − 918.9𝑥1 =-398.86.
෣
• Then, Φ ෩ 𝐻𝑜𝑢𝑟𝑠 = Φ ෩ −398.86 = Φ(−0.3518)
෩
𝜎 1133.7
𝜕𝐸(𝐻𝑜𝑢𝑟𝑠)
= 𝛽෨2 Φ(−0.3518)
෫ = 73.29𝑥 (1 −
𝜕𝐸𝑑𝑢𝑐
0.6368) = 73.29𝑥0.3638 = 26.34
(See the Table z in page 742)
7/9/2023 23

23

Comparing Tobit with OLS estimates

7/9/2023 24

24

12
 7/9/2023

The selection bias: wage of married

women (two-step)
• The selection equation, for a latent variable (unobserved):
𝑧𝑖∗ = 𝛾1 + 𝛾2 𝑤𝑖 + 𝑢𝑖 , 𝑖 = 1, … , 𝑁 (15)
• The indicator variable:
1 𝑧𝑖∗ > 0
𝑧𝑖 = { (16)
0 𝑜𝑡ℎ𝑒𝑟𝑤𝑖𝑠𝑒
• The outcome equation:
𝑦𝑖 = 𝛽1 + 𝛽2 𝑥𝑖 + 𝑒𝑖 , 𝑖 = 1, … 𝑛; 𝑁 > 𝑛 (17)
𝑦𝑖 = 𝛽1 + 𝛽2 𝑥𝑖 + 𝛽λ λ෨ 𝑖 + 𝑣𝑖 , 𝑖 = 1, … 𝑛 (18)
φ(෥𝛾1 +෥𝛾2 𝑤𝑖 )
Where, the “invesre Mills ratio” λ෨ 𝑖 = (19)
Φ(෥
𝛾1 +෥
𝛾2 𝑤𝑖 )

7/9/2023 25

25

Addressing selection bias: Wages of

married women (two-step)

7/9/2023 26

26

13
 7/9/2023

Addressing selection bias: Wages of

married women (two-part, full model)

7/9/2023 27

27

Further issues with Heckman models

• Two-step or two-part models?
• Two-step: Heckman, J. J. (1979). Sample selection bias as
a specification error. Econometrica: Journal of the
econometric society, 153-161. #34419
• Two part: Manning, W. G., Duan, N., & Rogers, W. H.
(1987). Monte Carlo evidence on the choice between
sample selection and two-part models. Journal of
econometrics, 35(1), 59-82. #478
• Further reading: Heckman model on Stata.
• The IVs in the selection model.
• Testing selection bias? Taking logs of zero value.

7/9/2023 28

28

14
 7/9/2023

Further practice and extensions

• Further estimators for choice behavior:
• For binary choice (probit vs. logit)
• Multinomial logit (mlogit)
• Conditional logit (clogit ???)
• Ordered choice (oprobit vs. ologit)
• Count data (poisson)
• Results interpretation
• Marginal effects for Probit and Tobit, log-log and log-
linear model (again)
• Further on selection bias
• Endogeneity again ☺

7/9/2023 29

29

15