Probit and Logit Models

Outline
• Linear probability model
• Probit and logit models
• Maximum likelihood estimation
• Coefficients
• Predicted probabilities
• Marginal effects
• Marginal effect at the means
• Average marginal effect
• Goodness of fit measures
• Pseudo R-squared
• Percent correctly predicted
Binary dependent variable
• A binary dependent variable has two outcomes: 0 or 1.
• Examples: working or not working, has insurance or does not have
insurance, etc.
• The outcome of interest is denoted as 1.
• 𝑦𝑦 = 1 if working, 𝑦𝑦 = 0 if not working.
• If the outcome of not working is of interest, then it would be denoted
as 1.
• 𝑦𝑦 = 1 if not working, 𝑦𝑦 = 0 if working.
• There are typically fewer outcomes of interest, i.e. fewer 1s in the
data.
Linear probability model (LPM)
• A linear probability model is a linear regression model where the dependent
variable is a binary variable.
• Linear probability model with binary dependent variable 𝑦𝑦 = 0 or 1.
• 𝑦𝑦 = 𝛽𝛽0 + 𝛽𝛽1 𝑥𝑥1 + 𝛽𝛽2 𝑥𝑥2 +. . +𝛽𝛽𝑘𝑘 𝑥𝑥𝑘𝑘 + 𝑢𝑢 = 𝑥𝑥𝑥𝑥 + 𝑢𝑢
• where 𝑥𝑥𝑥𝑥 is expressed in a matrix form.
• Expected value of 𝑦𝑦 is 𝐸𝐸(𝑦𝑦) = 𝑥𝑥𝑥𝑥.
• Because the binary variable 𝑦𝑦 has two outcomes 0 or 1, the expected value for 𝑦𝑦
is the probability of 𝑦𝑦 being 1, 𝑃𝑃(𝑦𝑦 = 1).
• 𝐸𝐸(𝑦𝑦) = 1 ∗ 𝑃𝑃(𝑦𝑦 = 1) + 0 ∗ 𝑃𝑃(𝑦𝑦 = 0) = 𝑃𝑃(𝑦𝑦 = 1)
• Example: if 30% of 𝑦𝑦 are 1 and the rest are zero, then 𝐸𝐸 𝑦𝑦 = 𝑃𝑃 𝑦𝑦 = 1 = 0.3
• The linear probability model for the probability of the outcome 𝑦𝑦 = 1 is
𝑃𝑃(𝑦𝑦 = 1) = 𝑥𝑥𝛽𝛽
Advantages and disadvantages of LPM
• Advantages of LPM
• Easy to estimate and interpret (coefficients are marginal effects)
• The coefficients and predictions are reasonably good
• Disadvantages of LPM
• Not the best model for binary dependent variable (probit or logit models are
better)
• Predicted probabilities can be less than 0 or greater than 1
• Marginal effects are the coefficients, which are constant/do not vary with 𝑥𝑥
• Heteroscedasticity because the variance is not constant
• 𝑣𝑣𝑣𝑣𝑣𝑣 𝑦𝑦 = 𝑃𝑃 𝑦𝑦 = 1 ∗ [1 − 𝑃𝑃(𝑦𝑦 = 1)]
Linear versus non-linear probability models
• The linear probability model estimate the probability of 𝑦𝑦 = 1 as a
linear function of the independent variables.
• 𝑃𝑃 𝑦𝑦 = 1 = 𝛽𝛽0 + 𝛽𝛽1 𝑥𝑥1 + 𝛽𝛽2 𝑥𝑥2 +. . +𝛽𝛽𝑘𝑘 𝑥𝑥𝑘𝑘 = 𝑥𝑥𝑥𝑥
• The probit and logit models estimate the probability of 𝑦𝑦 = 1 as a
non-linear function 𝐺𝐺 of the independent variables.
• 𝑃𝑃 𝑦𝑦 = 1 = 𝐺𝐺 𝛽𝛽0 + 𝛽𝛽1 𝑥𝑥1 + 𝛽𝛽2 𝑥𝑥2 +. . +𝛽𝛽𝑘𝑘 𝑥𝑥𝑘𝑘 = 𝐺𝐺(𝑥𝑥𝑥𝑥)
• 𝐺𝐺 is a non-linear function that transforms 𝑥𝑥𝑥𝑥 to be between 0 and 1 because
𝑃𝑃(𝑦𝑦 = 1) is a probability.
Normal distribution – pdf and cdf
• The probability density function (pdf) of the normal distribution 𝜙𝜙
shows the probability that 𝑦𝑦 is between two numbers.
• The cumulative density function (cdf) of the normal distribution Φ
shows the probability that 𝑦𝑦 is less than a given number.
pdf cdf
0.45 1.2
0.4
1
0.35
0.8
0.3
0.25 0.6
0.2
0.4
0.15
0.2
0.1
0.05 0
-2.5 -2 -1.5 -1 -0.5 0 0.5 1 1.5 2 2.5
0
-2.5 -2 -1.5 -1 -0.5 0 0.5 1 1.5 2 2.5
Probit model
• The probit model uses the cumulative density function (cdf) of the
normal distribution Φ.
𝑥𝑥𝑥𝑥
• 𝑃𝑃 𝑦𝑦 = 1 = Φ 𝑥𝑥𝑥𝑥 = ∫−∞ 𝜙𝜙 𝑧𝑧 𝑑𝑑𝑑𝑑
• 𝑃𝑃(𝑦𝑦 = 1) will be a number between 0 and 1 because the cdf of the
normal distribution is a number between 0 and 1.
Logit model
• The logit model uses the logistic function:
exp 𝑥𝑥𝑥𝑥 𝑒𝑒 𝑥𝑥𝑥𝑥
• 𝑃𝑃 𝑦𝑦 = 1 = 𝐺𝐺 𝑥𝑥𝑥𝑥 = =
1+exp 𝑥𝑥𝑥𝑥 1+𝑒𝑒 𝑥𝑥𝑥𝑥
• 𝑃𝑃(𝑦𝑦 = 1) will be a number between 0 and 1 because exp(𝑥𝑥𝑥𝑥) is
positive.
• The probability of 𝑦𝑦 = 0 is:
exp 𝑥𝑥𝑥𝑥 1
• 𝑃𝑃 𝑦𝑦 = 0 = 1 − 𝑃𝑃 𝑦𝑦 = 1 = 1 − =
1+exp 𝑥𝑥𝑥𝑥 1+exp 𝑥𝑥𝑥𝑥
Likelihood function
• The likelihood is the probability that the outcome for observation 𝑖𝑖 is
𝑦𝑦𝑖𝑖 .
• The likelihood of 𝑦𝑦𝑖𝑖 = 1 is 𝑃𝑃 𝑦𝑦𝑖𝑖 = 1 .
• The likelihood of 𝑦𝑦𝑖𝑖 = 0 is 𝑃𝑃 𝑦𝑦𝑖𝑖 = 0 .
𝑦𝑦𝑖𝑖 1−𝑦𝑦𝑖𝑖
• The likelihood function is defined as: 𝑃𝑃 𝑦𝑦𝑖𝑖 = 1 𝑃𝑃 𝑦𝑦𝑖𝑖 = 0
• The likelihood of 𝑦𝑦𝑖𝑖 = 1 is 𝑃𝑃 𝑦𝑦𝑖𝑖 = 1 1 𝑃𝑃 𝑦𝑦𝑖𝑖 = 0 1−1 = 𝑃𝑃 𝑦𝑦𝑖𝑖 = 1
• The likelihood of 𝑦𝑦𝑖𝑖 = 0 is 𝑃𝑃 𝑦𝑦𝑖𝑖 = 1 0 𝑃𝑃 𝑦𝑦 = 0 1−0 = 𝑃𝑃 𝑦𝑦 = 0
𝑖𝑖 𝑖𝑖
Maximum likelihood estimation
• The likelihood function is: 𝑃𝑃 𝑦𝑦𝑖𝑖 = 1 𝑦𝑦𝑖𝑖 𝑃𝑃 𝑦𝑦𝑖𝑖 = 0 1−𝑦𝑦𝑖𝑖
• Taking logs and summing up over all observations 𝑖𝑖.
• The log likelihood function is:
• ∑𝑛𝑛𝑖𝑖=1( 𝑦𝑦𝑖𝑖 ∗ 𝑙𝑙𝑙𝑙𝑙𝑙𝑙𝑙 𝑦𝑦𝑖𝑖 = 1 + 1 − 𝑦𝑦𝑖𝑖 ∗ 𝑙𝑙𝑙𝑙𝑙𝑙𝑙𝑙 𝑦𝑦𝑖𝑖 = 0 )
• Substituting 𝑃𝑃(𝑦𝑦 = 1) = 𝐺𝐺(𝑥𝑥𝑥𝑥) into the log likelihood function.
• ∑𝑛𝑛𝑖𝑖=1( 𝑦𝑦𝑖𝑖 ∗ log(𝐺𝐺 𝑥𝑥𝑥𝑥 ) + 1 − 𝑦𝑦𝑖𝑖 ∗ log(1 − 𝐺𝐺 𝑥𝑥𝑥𝑥 )
• The 𝛽𝛽 coefficients are obtained by maximizing the log likelihood
function.
Maximum likelihood estimation
• The probit and logit model coefficients are obtained by maximizing
the log likelihood function.
• max ∑𝑛𝑛𝑖𝑖=1( 𝑦𝑦𝑖𝑖 ∗ 𝑙𝑙𝑙𝑙𝑙𝑙𝑙𝑙 𝑦𝑦𝑖𝑖 = 1 + 1 − 𝑦𝑦𝑖𝑖 ∗ 𝑙𝑙𝑙𝑙𝑙𝑙𝑙𝑙 𝑦𝑦𝑖𝑖 = 0 )
• If the outcome 𝑦𝑦𝑖𝑖 = 1, the predicted probability 𝑃𝑃 𝑦𝑦𝑖𝑖 = 1 is maximized (e.g.
0.8 or 0.9).
• If the outcome 𝑦𝑦𝑖𝑖 = 0, 𝑃𝑃 𝑦𝑦𝑖𝑖 = 0 is maximized or equivalently the predicted
probability 𝑃𝑃 𝑦𝑦𝑖𝑖 = 1 is minimized (e.g. 0.1 or 0.2).
• The maximum likelihood estimators are consistent, asymptotically
normal, and asymptotically efficient if the assumptions hold.
Maximum likelihood estimation versus OLS
estimation
• The probit and logit model coefficients are obtained by maximizing
the log likelihood function (if the outcome 𝑦𝑦 = 1, the predicted
probability 𝑃𝑃 𝑦𝑦 = 1 is maximized)
• max ∑𝑛𝑛𝑖𝑖=1( 𝑦𝑦𝑖𝑖 ∗ 𝑙𝑙𝑙𝑙𝑙𝑙𝑙𝑙 𝑦𝑦𝑖𝑖 = 1 + 1 − 𝑦𝑦𝑖𝑖 ∗ 𝑙𝑙𝑙𝑙𝑙𝑙𝑙𝑙 𝑦𝑦𝑖𝑖 = 0 )

• The OLS coefficients are obtained by minimizing the sum of squared

residuals (difference between actual value 𝑦𝑦 and predicted values 𝑦𝑦)
�
• min ∑𝑛𝑛𝑖𝑖=1 𝑢𝑢� 2 = ∑𝑛𝑛𝑖𝑖=1(𝑦𝑦 − 𝑦𝑦) ̂ 2
� 2 = ∑𝑛𝑛𝑖𝑖=1(𝑦𝑦 − 𝑥𝑥𝛽𝛽)
Example
• Model to explain if women are in the labor force or not.
• 𝑃𝑃(𝑖𝑖𝑖𝑖𝑖𝑖𝑖𝑖 = 1) = 𝐺𝐺(𝛽𝛽0 + 𝛽𝛽1 𝑛𝑛𝑛𝑛𝑛𝑛𝑛𝑛𝑛𝑛𝑛𝑛𝑛𝑛𝑛𝑛 + 𝛽𝛽2 𝑒𝑒𝑒𝑒𝑒𝑒𝑒𝑒 + 𝛽𝛽3 𝑒𝑒𝑒𝑒𝑒𝑒𝑒𝑒𝑒𝑒 + 𝛽𝛽4 𝑎𝑎𝑎𝑎𝑎𝑎 + 𝛽𝛽5 𝑘𝑘𝑘𝑘𝑘𝑘𝑘𝑘𝑘𝑘𝑘𝑘𝑘)
• 𝑖𝑖𝑖𝑖𝑖𝑖𝑖𝑖 is a binary 0 or 1 variable for whether women are in labor force or not.
• 𝑛𝑛𝑛𝑛𝑛𝑛𝑛𝑛𝑛𝑛𝑛𝑛𝑛𝑛𝑛𝑛 is non-wife income.
• 𝑘𝑘𝑘𝑘𝑘𝑘𝑘𝑘𝑘𝑘𝑘𝑘𝑘 is number of kids under 6 years old.
• 57% of the women are in the labor force and the rest are not. The unconditional
probability of being in the labor force is 0.57. 𝑃𝑃 𝑦𝑦 = 1 = 0.57
Variable Mean Std. Dev. Min Max
inlf 0.57 0.50 0 1
nwifeinc 20.13 11.63 -0.03 96
educ 12.29 2.28 5 17
exper 10.63 8.07 0 45
age 42.54 8.07 30 60
kidslt6 0.24 0.52 0 3
LPM, probit, and logit model – coefficients
LPM Probit Logit • The coefficients are different for
VARIABLES inlf inlf inlf the probit and logit models. The
nwifeinc -0.003** -0.011** -0.020** logit coefficients are about 1.6
(0.001) (0.005) (0.008) times the probit coefficients.
educ 0.039*** 0.132*** 0.223*** • Interpretation of the coefficient on
(0.007) (0.025) (0.043) education: women with higher
exper 0.022*** 0.069*** 0.118*** education are more likely to be in
(0.002) (0.007) (0.013) the labor force.
age -0.019*** -0.058*** -0.095*** • Interpretation of the coefficient on
(0.002) (0.008) (0.013) age: women who are older are less
kidslt6 -0.275*** -0.886*** -1.464*** likely to be in the labor force.
(0.033) (0.117) (0.200) • The magnitudes of the coefficients
Constant 0.770*** 0.765* 1.153 are not interpreted.
(0.135) (0.440) (0.742)
Predicted probabilities
• After estimating the models and obtaining the coefficients 𝛽𝛽, ̂ the
predicted probabilities can be calculated as:
• 𝑃𝑃 𝑦𝑦𝑖𝑖 = 1 = 𝐺𝐺 𝛽𝛽̂0 + 𝛽𝛽̂1 𝑥𝑥𝑖𝑖1 + 𝛽𝛽̂2 𝑥𝑥𝑖𝑖2 +. . +𝛽𝛽̂𝑘𝑘 𝑥𝑥𝑖𝑖𝑘𝑘 = 𝐺𝐺(𝑥𝑥𝑖𝑖 𝛽𝛽)
̂
• If the actual value 𝑦𝑦𝑖𝑖 = 1 and the predicted probability 𝑃𝑃(𝑦𝑦𝑖𝑖 = 1) is above
0.5, it is a correct prediction.
• If the actual value 𝑦𝑦𝑖𝑖 = 0 and the predicted probability 𝑃𝑃(𝑦𝑦𝑖𝑖 = 1) is below
0.5, it is also a correct prediction.
• Otherwise, it will be incorrect prediction.
• The average of the predicted probabilities will be the unconditional
probability, which is the sample average 𝑦𝑦.
�
Actual values and predicted probabilities
LPM predicted Probit predicted Logit predicted
Actual value probability probability probability
obs i inlf y Inlfhat_lpm Inlfhat_probit Inlfhat_logit
1 1 0.65 0.67 0.68
2 1 0.73 0.77 0.77
3 1 0.61 0.63 0.64
4 1 0.72 0.76 0.76
601 0 0.31 0.27 0.26
602 0 0.35 0.31 0.29
603 0 0.39 0.35 0.34
604 0 0.67 0.69 0.70
605 0 -0.19 0.01 0.03
For the LPM, the predicted probabilities can be a negative number (e.g. -0.19 for observation 605) or a number
above 1, which is not reasonable.
For the probit and logit models, the predicted probabilities for the first four observations with inlf=1 are all
above 0.5 (correct prediction).
The predicted probabilities for observation 604 with inlf=0 are above 0.5 which in an incorrect prediction.
The predicted probabilities for the other four observations with inlf=0 are below 0.5 (correct prediction).
The average of inlf is 0.57, which is also the average of the predicted probabilities.
Marginal effects in the linear probability
model
• The linear probability model:
• 𝑃𝑃 𝑦𝑦 = 1 = 𝛽𝛽0 + 𝛽𝛽1 𝑥𝑥1 + 𝛽𝛽2 𝑥𝑥2 +. . +𝛽𝛽𝑘𝑘 𝑥𝑥𝑘𝑘 = 𝑥𝑥𝑥𝑥
• The coefficient on 𝑥𝑥𝑗𝑗 is 𝛽𝛽𝑗𝑗 .
• The marginal effect of 𝑥𝑥𝑗𝑗 on the probability of 𝑦𝑦 = 1 is the coefficient 𝛽𝛽𝑗𝑗 .
∆𝑃𝑃(𝑦𝑦 = 1)
= 𝛽𝛽𝑗𝑗
∆𝑥𝑥𝑗𝑗
• In the LPM, the marginal effects are the coefficients.
• The marginal effect explains the effect of the independent variable on the
probability that 𝑦𝑦 = 1.
Marginal effects in the probit and logit model
• The probit and logit model:
• 𝑃𝑃 𝑦𝑦 = 1 = 𝐺𝐺 𝛽𝛽0 + 𝛽𝛽1 𝑥𝑥1 + 𝛽𝛽2 𝑥𝑥2 +. . +𝛽𝛽𝑘𝑘 𝑥𝑥𝑘𝑘 = 𝐺𝐺(𝑥𝑥𝑥𝑥)
• The coefficient on 𝑥𝑥𝑗𝑗 is 𝛽𝛽𝑗𝑗 .
• The marginal effect of 𝑥𝑥𝑗𝑗 on the probability of 𝑦𝑦 = 1 is
∆𝑃𝑃(𝑦𝑦 = 1)
= 𝐺𝐺′ 𝑥𝑥𝑥𝑥 ∗ 𝛽𝛽𝑗𝑗
∆𝑥𝑥𝑗𝑗
• In the probit and logit model, the marginal effects are the coefficients multiplied
by a scale factor 𝐺𝐺 ′ 𝑥𝑥𝑥𝑥 , which is the derivative of the 𝐺𝐺 function.
• The marginal effect explains the effect of the independent variable on the
probability that 𝑦𝑦 = 1 (by how much the probability of 𝑦𝑦 = 1 increases when 𝑥𝑥𝑗𝑗
increases by 1 unit).
Marginal effects in the probit and logit model
• The marginal effect of 𝑥𝑥𝑗𝑗 on the probability of 𝑦𝑦 = 1 is
∆𝑃𝑃(𝑦𝑦=1)
• = 𝐺𝐺′ 𝑥𝑥𝑥𝑥 ∗ 𝛽𝛽𝑗𝑗
∆𝑥𝑥𝑗𝑗
• The probit model: 𝑃𝑃 𝑦𝑦 = 1 = Φ 𝑥𝑥𝑥𝑥
∆𝑃𝑃(𝑦𝑦=1)
• The marginal effect in the probit model: = 𝜙𝜙 𝑥𝑥𝑥𝑥 ∗ 𝛽𝛽𝑗𝑗
∆𝑥𝑥𝑗𝑗
• Φ is the cdf and 𝜙𝜙 is the pdf of the normal distribution.
exp 𝑥𝑥𝑥𝑥
• The logit model: 𝑃𝑃 𝑦𝑦 = 1 =
1+exp 𝑥𝑥𝑥𝑥
∆𝑃𝑃(𝑦𝑦=1) exp 𝑥𝑥𝑥𝑥
• The marginal effect in the logit model: = ∗ 𝛽𝛽𝑗𝑗
∆𝑥𝑥𝑗𝑗 [1+exp 𝑥𝑥𝑥𝑥 ]2
Marginal effect at the mean and
average marginal effect
∆𝑃𝑃(𝑦𝑦=1)
• The marginal effect depends on 𝑥𝑥. = 𝐺𝐺′ 𝑥𝑥𝑥𝑥 ∗ 𝛽𝛽𝑗𝑗
∆𝑥𝑥𝑗𝑗
• The marginal effect at the mean is calculated at the mean value of 𝑥𝑥, which is 𝑥𝑥.̅
∆𝑃𝑃(𝑦𝑦=1)
• = 𝐺𝐺′ 𝑥𝑥𝛽𝛽
̅ ∗ 𝛽𝛽𝑗𝑗
∆𝑥𝑥𝑗𝑗
• The average marginal effect is calculated for each observation, and then averaged
across all observations.
∆𝑃𝑃(𝑦𝑦=1) 1 𝑛𝑛
• = 𝐺𝐺𝐺 𝑥𝑥𝑖𝑖 𝛽𝛽 ∗ 𝛽𝛽𝑗𝑗 = ∑𝑖𝑖=1 𝐺𝐺𝐺 𝑥𝑥𝑖𝑖 𝛽𝛽 ∗ 𝛽𝛽𝑗𝑗
∆𝑥𝑥𝑗𝑗 𝑛𝑛
• The marginal effect at the mean uses the means of the variables, but there may
not be such “average” individual (e.g. mean for variable female is 0.3). The
average marginal effect makes more sense. In practice, the marginal effects will
be similar.
Marginal effect for an indicator variable
• If the model is: 𝑃𝑃 𝑦𝑦 = 1 = 𝐺𝐺(𝛽𝛽0 + 𝛽𝛽1 ∗ 𝑑𝑑1 + 𝛽𝛽2 𝑥𝑥2 )
• with an independent variable 𝑑𝑑1 which an indicator variable taking
values of 0 or 1, the marginal effect is calculated as:
• 𝐺𝐺 𝛽𝛽0 + 𝛽𝛽1 ∗ 1 + 𝛽𝛽2 𝑥𝑥2 − 𝐺𝐺 𝛽𝛽0 + 𝛽𝛽1 ∗ 0 + 𝛽𝛽2 𝑥𝑥2
• The marginal effect of 𝑑𝑑1 being 1 instead of 0 is the difference in
𝑃𝑃(𝑦𝑦 = 1) if 𝑑𝑑1 = 1 and the probability of 𝑃𝑃(𝑦𝑦 = 1) if 𝑑𝑑1 = 0.
Marginal effects
Probit marginal Probit average Logit marginal Logit average
effect at mean marginal effects effects at mean marginal effects
VARIABLES inlf inlf inlf inlf
nwifeinc -0.004** -0.003** -0.005** -0.004**
(0.002) (0.001) (0.002) (0.001)
educ 0.051*** 0.040*** 0.054*** 0.040***
(0.010) (0.007) (0.010) (0.007)
exper 0.027*** 0.021*** 0.029*** 0.021***
(0.003) (0.002) (0.003) (0.002)
age -0.023*** -0.018*** -0.023*** -0.017***
(0.003) (0.002) (0.003) (0.002)
kidslt6 -0.346*** -0.271*** -0.355*** -0.266***
(0.046) (0.031) (0.049) (0.031)
Unlike the coefficients, the marginal effects in the probit and logit model are similar. The marginal effects at the
mean and the average marginal effects are similar. The magnitude of the marginal effects can be interpreted.
For each additional year of education, women are 5.1% more likely to be in the labor force.
For each additional child less than 6 years old, women are 34.6% less likely to be in the labor force.
Pseudo R-squared
• Pseudo R-squared, aka McFadden R-squared, measures the goodness of fit for a probit or
logit model. It compares the log-likelihood of a model with that of a model with only a
constant.
2 𝐿𝐿𝐿𝐿𝑢𝑢𝑢𝑢
• 𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃 𝑅𝑅 = 1 −
𝐿𝐿𝐿𝐿0
• 𝐿𝐿𝐿𝐿𝑢𝑢𝑢𝑢 is the log likelihood for the unrestricted model with all independent variables.
• 𝐿𝐿𝐿𝐿0 is the log likelihood for the restricted model with only a constant.
• If the independent variables do not explain the dependent variable then the log likelihoods for the
restricted and unrestricted models (𝐿𝐿𝐿𝐿0 and 𝐿𝐿𝐿𝐿𝑢𝑢𝑢𝑢 ) will be the similar and the pseudo R-squared
will be 0.
• If the independent variables explain the dependent variable very well, then because the log
likelihood for the unrestricted model 𝐿𝐿𝐿𝐿𝑢𝑢𝑢𝑢 will be maximized (which is negative number that will
approach 0) and the pseudo R-squared will approach 1.
• The pseudo R-squared indicates how well the model predicts the outcome and how well
the model improves on a null model with only an intercept, but the magnitude is not
interpreted.
• A higher pseudo R-squared with the same dependent variable but different independent
variables would indicate that the model predicts the outcome better.
Pseudo R-squared example
2 𝐿𝐿𝐿𝐿𝑢𝑢𝑢𝑢 −406.30
• 𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃 𝑅𝑅 = 1 − =1− = 0.21
𝐿𝐿𝐿𝐿0 −514.87
• 𝐿𝐿𝐿𝐿𝑢𝑢𝑢𝑢 is the log likelihood for the unrestricted model with all independent
variables
• 𝐿𝐿𝐿𝐿0 is the log likelihood for the restricted model with only a constant.
• The pseudo R-squared shows how well the model predicts the
outcome, with a higher pseudo R-squared being preferred.
Review questions
• Describe and give examples of a binary dependent variable.
• Explain the functional form for the linear probability model, the
probit model, and the logit model.
• How are the coefficients interpreted?
• Describe the marginal effect at the mean and the average marginal
effect.
• How are the marginal effects interpreted?
• Describe the goodness of fit measures for the probit and logit model.