Scribd
Upload
364 views11 pages

Logit and Probit Models

Logit and probit models are used as alternatives to linear probability models to predict dichotomous outcomes while ensuring predicted probabilities remain between 0 and 1. They do this using the logistic and normal cumulative distribution functions respectively. Both can be derived from an underlying latent variable model. Maximum likelihood estimation is used instead of ordinary least squares due to the nonlinear nature of the models. While coefficients from logit, probit and LPM are not directly comparable, these models typically produce consistent results in terms of coefficient signs and significance. The marginal effects in logit and probit may diminish with different variable values compared to the constant effect in LPM.

Uploaded by

galih visnhu
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
364 views11 pages

Logit and Probit Models

Logit and probit models are used as alternatives to linear probability models to predict dichotomous outcomes while ensuring predicted probabilities remain between 0 and 1. They do this using the logistic and normal cumulative distribution functions respectively. Both can be derived from an underlying latent variable model. Maximum likelihood estimation is used instead of ordinary least squares due to the nonlinear nature of the models. While coefficients from logit, probit and LPM are not directly comparable, these models typically produce consistent results in terms of coefficient signs and significance. The marginal effects in logit and probit may diminish with different variable values compared to the constant effect in LPM.

Uploaded by

galih visnhu
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/ 11

LOGIT AND PROBIT MODELS

Diny Ghuzini
February 2009
Logit and Probit models are used to avoid the limitations
of LPM

Consider a model:
P( y  1 | x)  G( 0 1 x1...  k xk )  G(0  x )
0<G(z)<1

Functions suggested to make sure the probability between 0


and 1 are:
1. Logistic function
ez
G( z)  (cdf for standard logistic function)
1  ez
Logit model
2. Normal function
z
G ( z )  ( z )    (v)dv (cdf for standard normal function)


1 z2
 
 ( z )  (2 ) e 2 2

1
CDF

The function
increases quickly at
z=0

z
-∞ 0 ∞
Logit and Probit model can be derived from latent variable
model. Let y*: latent variable
y*  0  x  e, y  1[ y*  0]
P( y  1 | x)  P( y*  0 | x)  P[e  (0  x ) | x]
 1  G[(0  x )]  G(0  x )
Direction of xj effect on E(y*|x)=β0+xβ and on
E(y|x)=P(y=1|x)=G(β0+xβ) is always the same
But βj are not useful (refer to latent, unobservable variable)

To find partial effect:


p( x) dG
 g ( 0  x )  j where g ( z)  ( z)
x j dz
If x1 binary variable, partial effect from changing x1 from 0
to 1:
G(0  1   2 x2  ...   k xk )  G(0   2 x2  ...   k xk )

If xk discrete variable (such as number of children), partial


effect from changing xk from ck to ck +1 :
G[0  1x1  ...   k (ck  1)]  G(0  1x1  ...   k ck )
Method of estimation:
Because the function used in logit and probit model are
nonlinear, we cannot use OLS as in LPM, use maximum
likelihood estimation (MLE) instead.
This method will already account for heteroscedasticity.

To derive ML estimator:
density of yi given xi f ( y | xi ;  )  [G ( x i  )] y [1  G ( xi  )]1 y , y  0,1

Log-likelihood function ( )  yi log[ G( xi  )]  (1  yi ) log[ 1  G( xi  )]

MLE of β is derived by maximizing sum of log-likelihood for


all sample n

 (  )
i 1
If G(.) is standard logit cdf, then β is logit estimator
If G(.) is standard normal cdf, then β is probit estimator

After the standard error of each β is provided, we can


construct t test and confidence intervals as in OLS

Test multiple restrictions:


- LM test
- Wald test
- Likelihood ratio test
Estimating the effects of xj on P(y=1|x):
Pˆ ( y  1 | x)  [ g (ˆ 0 xˆ )ˆ j ]x j

often average x is used to calculate g

Logit, Probit and LPM estimates are not directly


comparable.
To compare Logit with Probit estimates:
- Multiply probit estimate by 1.6
- Multiply logit estimates by 0.625

To compare LPM with Logit and Probit estimates:


- Divide logit estimates by 4
- Divide probit estimates by 2.5
Independent variables LPM (OLS) Logit (MLE) Probit (MLE)
nwifeinc -0.0034 -0.021 -0.012
-2.267 -2.625 -2.4
educ 0.038 0.221 0.131
5.429 5.140 5.24
exper 0.039 0.206 0.123
6.5 6.4375 6.474
exper2 -0.0006 -0.0032 -0.0019
-3.333 -3.2 -3.167
age -0.016 -0.088 -0.053
-8 -5.867 -6.625
kidslt6 -0.262 -1.443 -0.868
-8.188 -7.074 -7.294
kidsge6 0.013 0.06 0.036
1 0.8 0.837
constant 0.586 0.425 0.27
3.881 0.494 0.530
Example: labor force of married woman in 1975
Results:
• Three models show consistent coefficient signs
• The same variables are stat. significant
• Coefficient magnitude not comparable

In Logit and Probit model, marginal effects not constant but


diminish
• LPM, coef kidslt6 = -0.262
one more small child reduces prob. In LF by 0.262
regardless the number of children
• Probit,
a women with nwifeinc = 20.13, educ = 12.3, exper = 10.6,
age = 42.5 (roughly sample average) and kidsge = 1 and
plug to (ˆ  ˆ x  ...  ˆ x ) Φ is CDF standard normal
0 1 1 k k

Probability of working if the woman have zero to one


small child is 0.373-0.707 = -0.334

Probability of working if the woman have one to two


small children is 0.117-0.373 = -0.256

The marginal effect of an increase in number of small


children diminishes.

You might also like