4.

Tobit-Model

Tobit-Model

1
University of Freiburg WS 2007/2008
Alexander Spermann
 4.
Tobit-Model 1. The Tobit - Model
2. An example

Wooldridge (2003), Introductory

Econometrics, 2nd edition, Chap.17.2

Other:
Wooldridge (2002): Econometric Analysis of Cross Section and
Panel Data, Chapter 16.
Ruud (2000): An Introduction to Classical Econometric Theory,
Chapter 28.
Greene (2000): Econometric Analysis, 4th edition, Chapter 20.3
2
University of Freiburg WS 2007/2008
Alexander Spermann
 4. • Problem:
Tobit-Model
special attribute(s) of the dependent
variable (DV)
1. dependent variable constrained and
2. clustering of observations at the constraint

• Examples:
• consumption (1. not 2.)
• wage changes (2. not 1.)
• Labor supply (1. and 2.)
3
University of Freiburg WS 2007/2008
Alexander Spermann
 4.
left- and right-censoring in the data
Tobit-Model
left-censored, from below right-censored, top-coded

Distribution of hourly Distribution of log-wages

benefits, Fringe.dta, in West-Germany, males,
1.1.1986, clerks, IABS01
4 command: hist hrbens
University of Freiburg WS 2007/2008
Alexander Spermann
 4.
Tobit-Model 2 different sorts of Models

• Data censoring
– Earnings variable (IABS)
– Demand for stadium tickets
– Duration in unemployment

• Corner solutions
– Labor Supply
– Household expenditures on holidays

5
University of Freiburg WS 2007/2008
Alexander Spermann
 4. Censoring in a regression framework
Tobit-Model
Ruud, Figure 28.2

6
University of Freiburg WS 2007/2008
Alexander Spermann
 4. • If DV is constrained and if there is clustering
Tobit-Model
– OLS on the complete sample biased
and inconsistent,
– OLS on the unclustered part biased
and inconsistent.

7
University of Freiburg WS 2007/2008
Alexander Spermann
 4.
Tobit-Model Solution possibility 1: Estimate a Probit Model

1 if y0
y
0
Loses information on y.

 y  R  do not throw
y away
 0 information
(Tobin 1958)
8 Solution: Tobit-regression
University of Freiburg WS 2007/2008
Alexander Spermann
 4.
*
Tobit-Model y
Trick: introduce a latent variable i
Assume: linear conditional expectation for latent Var.

E ( y | x)  x 
*
i
'
i
Assumption:
y  x   i
*
i
'
i i ~ i.i.d. N(0,  ) 2

y *
if y 0
*
i
yi   i

0 if yi*  0
9
University of Freiburg WS 2007/2008
Alexander Spermann
 4.
Tobit-Model Random sample {(x i , yi ) : i  1, 2,..., N}

Estimation of the parameters of the model:

• Non-linear LS estimation
• Maximum likelihood method

10
University of Freiburg WS 2007/2008
Alexander Spermann
 4. Maximum likelihood estimation:
Tobit-Model

Likelihood-function consists in two parts

1. Probit-Part
For censored observations we have:
Pr(yi  0)  Pr(y*i  0)
 i x i'  
 Pr(i   x )  Pr   
'
i 
   
 x i'    x i'  
    1   
11      
University of Freiburg WS 2007/2008
Alexander Spermann
 4.
2. Linear part
Tobit-Model
Can formulate a linear model for the part
that is uncensored:
lim Pr(yi  Yi  y i   | yi  0,   0) 
0

  yi  x i'      yi  x i'   
lim       
0
      
1  yi  x i'  
f ( i )    
   

12
University of Freiburg WS 2007/2008
Alexander Spermann
 4. Likelihood- and Log-Likelihood-function:
Tobit-Model
  x i'    1  yi  x i'  
L   1        
yi  0     iy  0    
  x i'     1  yi  x i'   
ln L   ln 1        ln    
yi  0      yi  0     

• ln L is maximized wrt β and σ.

• FOC yields estimator for β and σ.
• β and σ are asymptotically normal.
13 Inference is standard.
University of Freiburg WS 2007/2008
Alexander Spermann
 4. Data:
Tobit-Model
Dependent Variable:
• hours working hours (yearly)
of married women
753 Observations
• 428 women exchange work for money
in the labor market
(hours vary in the dataset between 12
and 4950)
• 325 women do not work.
14
University of Freiburg WS 2007/2008
Alexander Spermann
 4. explanatory variables:
Tobit-Model
• age age

• educ education
in years of schooling

• exper experience
in actual years of work

• nwifeinc family income (in 1000$) that is not

generated by the woman

• kidslt6 number of kids

age < 6

• kidsge6 number of kids

6< age < 18
15
University of Freiburg WS 2007/2008
Alexander Spermann
 4. Estimation of a Tobit-Model (in Stata):
Tobit-Model

Source: Wooldridge, Econometric Analysis of Cross Section and Panel Data (2002)

 estimated coefficients are to be interpreted as

16 the effect of the regressors on the latent variable.
University of Freiburg WS 2007/2008
Alexander Spermann
 4. Direct comparison of OLS and Tobit output
Tobit-Model impossible
OLS Tobit
nwifeinc -3.45 -8.81
Dependent
educ 28.76 80.65 variable:
exper 65.67 131.56 hours
exper2 -0.700 -1.86
age -30.61 -54.41
Kidslt 6 -442.09 -894.02
Kidsge 6 -32.78 -16.22
Constant 1330.48 965.31
Log- likelihood ---- -3819.09
R2 0.266 0.274
17 ˆ 750.18 1122.02
University of Freiburg WS 2007/2008
Alexander Spermann
 4.
Tobit-Model
1. Marginal effect on the latent variable

E  y* | x 
 k
xk

Slope of dashed
line: tobit

Slope of solid line:

OLS

18
University of Freiburg WS 2007/2008
Alexander Spermann
 4.
Tobit-Model
2. Marginal effect on the actual
variable
E  y | x   x 
 k  
xk   
Probability that an observation is different from
zero (if 1, then OLS=Tobit)

y
Green line!!

0
19 x
University of Freiburg WS 2007/2008
Alexander Spermann
 4. 3. Marginal effect on positive observations
Tobit-Model
E(y | x, y  0)   c 
 k  k  k {1   (c)[c   (c)]}  k
x k c

Where λ(c) is called inverse Mills Ratio:

 x 
 
(c) 
 (c)    
 (c)  x 
 
  

λ(c) captures the change in the population, we

condition on (y>0), when changing x.

20
University of Freiburg WS 2007/2008
Alexander Spermann
 4. 4. Marginal effect on the probability, that an
Tobit-Model observation is uncensored.

 x   x 
Pr(y  0 | x)  1         
     

It follows:

 Pr  y  0 | x   x  k
  
xk   

NB: For coefficients 2-4 need choose an appropriate

x-vector!
21
University of Freiburg WS 2007/2008
Alexander Spermann
 4. Comparison OLS - TOBIT on the basis of
Tobit-Model
the marginal effect on actual DV (example
educ, for an average individual):

OLS TOBIT
 x' 
̂ k ,OLS k  


 

80.65 0.604
28.76 48,73

22
University of Freiburg WS 2007/2008
Alexander Spermann
 4. Interpretation:
Tobit-Model
On average, an additional year of
education increases the labor supply by
48,7 hours (for an average individual).

 OLS underestimates the effect of

education on the labor supply (in the
average of the explanatory variables).

23
University of Freiburg WS 2007/2008
Alexander Spermann
 4. • dtobit calculates the four different
Tobit-Model
marginal effects (at the mean of the
explanatory variables):

24
University of Freiburg WS 2007/2008
Alexander Spermann
 4.
Tobit-Model

25
University of Freiburg WS 2007/2008
Alexander Spermann
 4.
Tobit-Model

26
University of Freiburg WS 2007/2008
Alexander Spermann
 4. Specification
Tobit-Model
• Unobserved, independent
heterogeneity → not problematic, as
OLS
• Endogeneity (left-out variables,
simultaneity) → „standard-IV“, similar
to OLS
• Heteroskedasticity, nonormal errors →
inconsistency, different from OLS

27
University of Freiburg WS 2007/2008
Alexander Spermann
 4.
Tobit-Model
alternatives for Tobit

• nonlinear estimation, eg. E(Y|

x)=exp(xb)
• CLAD-estimator
(for censoring problems)
• hurdle models, two-tiered models
(for corner solution problems)

28
University of Freiburg WS 2007/2008
Alexander Spermann