Chapter 2

CHAPTER 2
QUALITATIVE EXPLANATORY AND REGRESSAND VARIABLE

OF REGRESSION MODELS
8/1/2021 For MSc students by Urgaia R.(Ph.D.),OSU 1

2.1 Qualitative Explanatory Variables in Regression Models
 In regression analysis , in addition to quantitative variables, the

dependent variable may be essentially influenced by qualitative nature
of the variables.
 These are marital status, job category, season, gender, race, colour,
religion, nationality, geographical region, party affiliation, and political
upheavals and etc.
 These qualitative variables are essentially nominal scale variables,
having no particular numerical values, which are known as dummy
variables, indicator variables, categorical variables, and qualitative
variables.
 But we can "quantify" them by taking values of 0 and 1, where 0
indicating the absence of an attribute and 1 indicating its presence.
 Thus the gender variable can be quantified as male = 1 and female =0,
or vice versa and We use them in the regression equation together
with the other independent variables.

But random assignment can’t always control for all possible other
factors—though sometimes we may be able to identify some of these
factors and add them to our equation
Let’s say that the treatment is job training:
◦ Suppose that random assignment, by chance, results in one group
having more males and being slightly older than the other group
◦ If gender and age matter in determining earnings, then we can
control for the different composition of the two groups by including
gender and age in our regression equation:
OUTCOMEi = β0 + β1TREATMENTi + β2D1i + β3X2i + εi
where: D1i = dummy variable for the individual’s gender
X2 = the individual’s age

 Unfortunately, random assignment experiments are not common in
economics because they are subject to problems that typically do not
plague medical experiments—e.g.:
1. Non-Random Samples:
 Most subjects in economic experiments are volunteers, and samples
of volunteers often aren’t random and therefore may not be
representative of the overall population
 As a result, our conclusions may not apply to everyone
2. Unobservable Heterogeneity:
 In Equation the above, we added observable factors to the equation to
avoid omitted variable bias, but not all omitted factors in economics
are observable
 This “unobservable omitted variable” problem is called unobserved
heterogeneity

3. The Hawthorne Effect:
 Human subjects typically know that they’re being studied, and they
usually know whether they’re in the treatment group or the control
group
 The fact that human subjects know that they’re being observed
sometimes can change their behavior, and this change in behavior
could clearly change the results of the experiment
4. Impossible Experiments:
 It’s often impossible (or unethical) to run a random assignment
experiment in economics
 Think about how difficult it would be to use a random assignment
experiment to study the impact of marriage on earnings!

 2nd, If a qualitative variable has m categories, you may include m
dummies, provided you do not include the (common) intercept in the
model. This way we do not fall into the dummy variable trap.
 3rd, The category that the value of 0 is called the reference, benchmark
or comparison category. All comparisons are made in relation to the
reference category, as we will show with our example.
 4th , if there are several dummy variables, you must keep track of the
reference category; otherwise, it will be difficult to interpret the
results.
 5th, at times we will have to consider interactive dummies, which we
will illustrate shortly.
 6th, since dummy variables take values of 1 and 0, we cannot take their
logarithms. That is, we cannot introduce the dummy variables in log
form.

Table 2.1: Regression Results for a model of wage determination:
Dependent Variable: WAGE; Method: Least Squares, Sample: 1 1289
Included observations: 1289
Coefficient Std. Error t-Statistic Prob. Values
CONSTANT -7.183338 1.015788 -7.071691 0.0000
FEMALE -3.074875 0.364616 -8.433184 0.0000
NONWHITE -1.565313 0.509188 -3.074139 0.0022
UNION 1.095976 0.506078 2.165626 0.0305
EDUCATION 1.370301 0.065904 20.79231 0.0000
EXPER 0.166607 0.016048 lO.38205 0.0000
R-squared 0.323339 Mean dependent var 12.36585
Adjusted R-squared 0.320702 S.D. dependent var 7.896350
S.E. of regression 6.508137 Akaike info criterion 6.588627
Sum squared resid 54342.54 Schwarz criterion 6.612653
Log likelihood -4240.370 Durbin-Watson stat 1.897513
F-statistic 122.6149 Prob(F -statistic) 0.000000

Interactive dummies
 To find out the interactive dummies, we re-estimate the wage function above by adding
to it the product of the female and non-white dummies. Such a product is called an
interactive dummy, for it interacts the two qualitative variables. Adding the interactive
dummy, we obtain the results in reported in the following table.
Table 2.2: Wage function with interactive dummies.
Dependent Variable: WAGE, Method: Least Squares, Sample: 1 1289

Constant -7.0887 1.0195 -6.9533 0.0000
-3.2401 0.3953 -8.1961 0.0000
-2.1585 0.7484 -2.8840 0.0040
1.1150 0.5063 2.2021 0.0278
Education 1.3701 0.0659 20.791 0.0000
Experiences 0.1658 0.0161 10.326 0.0000
1.0954 1.0128 1.0814 0.2797

Table 2.3: Wage function with differential intercept and slope dummies.
Variable Coefficient Std. Error t-Statistic Prob. Values
Constant -11.09129 1.421846 -7.8033 0.0000
3.174158 1.966465 1.614144 0.1067
2.909129 2.780066 1.046424 0.2956
4.454212 2.973494 1.497972 0.1344
Edu 1.587125 0,093819 16.91682 0.0000
Exp 0.220912 0,025107 8.798919 0.0000
-0.336888 0,131993 -2,552314 0.Q108
-0.096125 0.031813 -3,021530 0.0026
-0.321855 0,195348 -1.647595 0.0997
-0.022041 0.044376 -0.496700 0.6195
-0.198323 0.191373 -1.036318 0.3003
-0.033454 0.046054 -0,726410 0,4677

2

Table 2.4 Reduced wage function.
Dependent Variable: WAGE, Method: Least Squares, Sample: 1
1289
Cons -10.645 1.3718 -7.7600 0.0000

FE 3.2574 1.9592 1.6626 0.0966
NW 2.6269 2.4178 1.0864 0.2775
UN 1.0785 0.5053 2.1339 0.0330
Edu 1.5658 0.0918 17.054 0.0000
Exp 0.2126 0.0227 9.3381 0.0000
FE*Edu -0.3469 0.1314 -2.6386 0.0084
FE*Exp -0.0949 0.0315 -3.0074 0.0027
NW*Edu -0.3293 0.1866 -1.7648 0.0778

Of course, there are other possibilities to express the wage function.
 For example, you may want to interact female with union and education
(female*union*education), which will show whether the females who are
educated and belong to unions have differential wages with respect to
education or union status.
 But beware of introducing too many dummy variables, for they can rapidly
consume degrees of freedom..
Table 2.5 Semi-log model of wages.
Dependent Variable: lnWAGE, Method: Least Squares,
Constant 0.9055 0.0741 12.2076 0.0000
-0.2491 0.0266 -9.3578 0.0000
-0.1335 0.0371 -3.5913 0.0003
0.1802 0.0369 4.8763 0.0330
Edu 0.0998 0.0048 20.752 0.0000
Exp 0.0127 0.0012 10.889 0.0000
1
Table 2.6 Regression of GPI on GPS with interactive dummy.
Dependent Variable: GPI, Method: Least Squares,
Sample: 1959-2007
Variable Coefficient Std. Error t-Statistic Prob.
Values
Constant -32.490 23.249 -1.3974 0.1691
1.0692 0.0259 41.256 0.0000
-327.85 61.753 -5.3089 0.0000
0.2441 0.0445 5.4747 0.0000

.
.
.

2.2 Limited Dependent Variable Models
 The standard linear regression models are applied when the behaviour of
economic agents approximated by continuous variables such as income,
saving, expenditure, output, etc.
 But there are many situations in which the dependent variable in a
regression equation simply represents a discrete choice assuming only a
limited number of values.
 Models involving dependent variables of this kind are called limited or
discrete dependent variable models also called qualitative response
models.
 Examples of discrete choices (a) whether to work or not, (b) whether to
attend formal education or not, (c) Choice of occupation and (d) Brand of
consumer durable goods to purchase.
 There are three approaches to developing a probability model for a binary
response variable: Linear Probability, Logit and Probit models.

Probability

1

2.2.1 of the Logit model
𝑛
𝑦𝑖 1−𝑦 𝑖
𝐿= 𝐺(𝑋𝑖 𝛽) 1 − 𝐺(𝑋𝑖 𝛽)
𝑖=1

Applying natural logarithms, we can get the log-likelihood function as:
𝑛
𝑙𝑛𝐿 = 𝑦𝑖 𝑙𝑛 𝐺(𝑋𝑖 𝛽) + (1 − 𝑦𝑖 )𝑙𝑛 1 − 𝐺(𝑋𝑖 𝛽)

𝑖=1
This equation is highly non-linear, and we need to apply numerical optimization

methods to obtain the solutions.

𝑛
𝑙𝑛𝐿 = 𝑦𝑖 𝑙𝑛 𝐺(𝑋𝑖 𝛽) + (1 − 𝑦𝑖 )𝑙𝑛 1 − 𝐺(𝑋𝑖 𝛽)

𝑖=1

𝑛
𝑙𝑛𝐿 = 𝑦𝑖 𝑙𝑛 𝑒 𝛼 +𝑋𝑖 𝛽 − 𝑦𝑖 ln⁡

(1 + 𝑒 𝛼 +𝑋𝑖 𝛽 − (1 − 𝑦𝑖 ) ln⁡
(1 + 𝑒 𝛼 +𝑋𝑖 𝛽
𝑖=1
The first order conditions are:

𝒏 𝒏 𝒏 𝒏
𝒀𝒊 − 𝜶 + 𝜷𝑿𝒊 𝑿𝒊 = 𝟎 ⇒ 𝒀𝒊 𝑿𝒊 = 𝜶 𝑿𝒊 + 𝜷 𝑿𝒊 𝟐
𝒊=𝟏 𝒊=𝟏 𝒊=𝟏 𝒊=𝟏
𝒏 𝒏 𝒏
𝒀𝒊 − 𝜶 + 𝜷𝑿𝒊 =𝟎⇒ 𝒀𝒊 = 𝒏𝜶 + 𝜷 𝑿𝒊
𝒊=𝟏 𝒊=𝟏 𝒊=𝟏

Odds Ratio : CDF
We realize that the sigmoid, or S-

shaped, curve in the figure is
very much resembles the
cumulative distribution function
(CDF) of a random variable.
Therefore, one can easily use the
CDF to model regressions where
the response variable is
dichotomous, taking 0–1 values.
Fig. A cumulative distribution function (CDF) can be both for logit

and probit models.
odds( x  1)  p ( x) 
 e  odds( x)  
odds( x)  1  p ( x) 

• Step 2: Raise e = 2.718 to the lower and upper bounds of the Confidence
Interval :
•If entire interval is above 1, conclude positive association

• If entire interval is below 1, conclude negative association
• If interval contains 1, cannot conclude there is an association

Multiple or Multinomial Logistic Regression
• Extension to more than one predictor variable either numeric or
dummy variables with k predictors, the model is written:
e  0  1x1   k xk i
G ( x,  )   0  1 x1   k xk
 e ( Adgusted)
1 e
• Adjusted Odds ratio for raising xi by 1 unit, holding all other predictors
constant.

• The logit model uses the logistic probability distribution to estimate
the parameters of the model.
• Although seemingly nonlinear, the log of the odds ratio makes the logit
model linear in the parameters.
• If we have grouped data, we can estimate the logit model by OLS
whereas in the cases of micro-level data, we have to use the method of
maximum likelihood.
• In the former case we will have to correct for heteroscedasticity in the
error term.
• Unlike the LPM, the marginal effect of a regressor in the logit model
depends not only on the coefficient of that regressor but also on the
values of all regressors in the model.
• Until the availability of readily accessible computer packages to estimate the logit
and probit models, the LPM was used quite extensively because of its simplicity.

• An alternative to logit is the probit model. The underlying probability
distribution of probit is the normal distribution.
• The parameters of the probit model are usually estimated by the
method of maximum likelihood.
• Like the logit model, the marginal effect of a regressor in the probit
model involves all the regressors in the modeL
• The logit and probit coefficients cannot be compared directly. But if you
multiply the probit coefficients by 1.81, they are then comparable with
the logit coefficients.
• This conversion is necessary because the underlying variances of the
logistic and normal distribution are different.
• In practice, the logit and probit models give similar results. The choice
between them depends on the availability of software and the ease of
interpretation.
Table 2.7 Linear probability model of to smoke or not to smoke.
Constant 1.1230 0.1883 5.9625 0.0000
Age -0.0047 0.0008 -5.7009 0.0000
Edu -0.0206 0.0046 -4.4652 0.0040
Income 1.03E-06 1.63E-06 0.6285 0.5298
Pcig 0.0051 0.0028 -0.7990 0.0723
R-squared 0.0387
Adjusted R-squared 0.0355

• We can refine this model by introducing interaction terms, such as age
multiplied by education, or education multiplied by income, or
introduce a squared term in education or in age to find out if there is
nonlinear impact of these variables on smoking.
• But there is no point in doing that, because the LPM has several
inherent limitations.
• Firstly, the LPM assumes that the probability of smoking moves linearly
with the value of the explanatory variable, no matter how small or
large that value is.
• Secondly, by logic, the probability value must lie between 0 and 1. But
there is no guarantee that the estimated probability values from the
LPM will lie within these limits. This is because OLS does not take into
account the restriction that the estimated probabilities must lie within
the bounds of 0 and 1.

Estimation of the logit model
 Estimation of the logit model depends on the type of data available for
analysis, either data at the individual, or micro, level, as in the case of the
smoker example, and data at the group level.
Table 2.8.:Logit model of to smoke or not to smoke.
Dependent Variable: SMOKER, Method: Maximum Likelihood(ML) - Binary Logit
(Quadratic hill climbing), Sample: 1 1196, Included observations: 1196
Convergence achieved after 3 iterations QML (Huber/White) standard errors &
covariance
Constant 2.7450 0.8217 3.3405 0.0008
Age -0.0208 0.0036 -5.7723 0.0000
Edu -0.0909 0.0205 -4.4274 0.0000
Income 4.72E-06 7.27E-06 0.6490 0.5163
Pcig -0.0223 0.0123 -1.8016 0.0716
McFadden R-squared = 0.0297, LR statistic= 47.26 with Prob(LR statistic)=0.000

• Let us examine these results. The variables age and education are highly
statistically significant and have the expected signs. As age of the people
increases, the value of the logit decreases, or they are less likely to smoke
perhaps due to health concerns that is as people age.
• Likewise, more educated people are less likely to smoke, perhaps due to
the ill effects of smoking.
• The price of cigarettes has the expected negative sign and is significant at
about the 7% level. Ceteris paribus, the higher the price of cigarettes, the
lower is the probability of smoking.
• Income has no statistically visible impact on smoking, perhaps because
expenditure on cigarettes may be a small proportion of family income.

 That is, the probability that a person with the given characteristics is a smoker is
about 38% and this person is a smoker.

• We can test the null hypothesis that all the coefficients are
simultaneously zero with the likelihood ratio (LR) statistic, which is
the equivalent of the F-test in the linear regression model.
• Under the null hypothesis that none of the regressors are significant,
the LR statistic follows the chi-square distribution with df equal to the
number of explanatory variables, which four.
• As you can see at the bottom of the above Table2.8 shows, the value of
the LR statistic is about 47.26 and the p value is practically zero, thus
refuting the null hypothesis.
• Therefore, the four variables included in the logit model are important
determinants of smoking habits .

Table 2.9: The logit model of smoking with interaction.
Dependent Variable: SMOKER Method: ML Binary Logit (Quadratic hill climbing),
Sample: 1 1196, Included observations: 1196, Convergence achieved after 10
iterations and Covariance matrix computed using second derivatives
Constant 1.0931 0.9556 1.1438 0.2527
Age -0.0182 0.0037 -4.8112 0.0000
Edu 0.0394 0.0425 0.9281 0.3533
Income 9.50E-05 2.69E-06 3.5351 0.0004
Pcig -0.0217 0.0125 -1.7324 0.0832
-7.45E-06 2.13E-06 -3.4897 0.0005
McFadden R-squared =0.0377, LR statistic= 59.96 with Prob(LR statistic)=0.000

• In Table 2.8, individually education had a significant negative impact on
the logit (probability of smoking) and income had no statistically
significant impact.
• However, introducing the multiplicative or interactive effect of the two
variables as an additional explanatory variable, the results are given in
Table 2.9.
• Now education by itself has no statistically significant impact on the logit,
but income has highly significant positive impact.
• But if you consider the interactive term, education multiplied by income,
it has significant negative impact on the logit.

• That is, persons with higher education who also have higher incomes
are less likely to be smokers than those who are more educated only or
have higher incomes only.
• What this suggests is that the impact of one variable on the probability
of smoking may be reinforced by the presence of other variable(s).
• In the LPM, the error term has non-normal distribution; in the logit
model, the error term has the logistic distribution and in the probit
model, it has the normal distribution.

Table 2.10: ML estimation of the censored regression model.
Dependent Variable: HOURS Method: ML - Censored Normal (TOBIT) (Quadratic
hill climbing) Sample: 1 753,Included observations: 753, Left censoring (value) at
zero, Convergence achieved after 6 iterations and Covariance matrix computed
using second derivatives
Coefficient Std. Error Z-Statistic Prob. Values
Constant 1126.3 379.58 2.9672 0.0030
AGE -54.109 6.6213 -8.1720 0.0000
EDU 38.646 20.684 1.8683 0.0617
EXP 129.82 16.229 7.9993 0.0000
EXPQ -1.8447 0.5096 -3.6194 0.0003
FAMINC 0.0407 0.5096 7.7540 0.0000
KIDSLT6 -782.37 103.75 -7.5408 0.0000
HUSWAGE -105.509 15.629 -6.7507 0.0000

Interpretation of the Tobit estimates
 For example, if the husband's wages go up, on average, a woman will
work less in the labour market, ceteris paribus.
 The slope coefficients of the various variables in Table 2.10 gives the
marginal impact of that variable on the mean value of the latent
variable
 But in practice we are interested in the marginal impact of a regressor
on the mean value of the actual values observed in the sample.
 In the Tobit type censored regression models, a unit change in the
value of a regressor has two effects:
(1) the effect on the mean value of the observed regressand, and
(2) the effect on the probability that the latent variable is actually
observed.

2.10

2.10

 As in the Tobit model, an individual regression coefficient in truncated
samples measures the marginal effect of that variable on the mean
value of the regressand for all observations - that is, including the non-
included observations.
 Hence, the within-sample marginal effect of a regressor is smaller in
absolute value than the value of the coefficient of that variable, as in the
case of the Tobit model.
 Assuming that the error term follows the normal distribution with zero
mean and constant variance, we can estimate censored regression
models by the method of maximum likelihood (ML). The estimators
thus obtained are consistent.
 Since the Tobit model uses more information or observations than the
truncated regression model, the estimates obtained from Tobit are
expected to be more efficient.

 One major caveat is that the ML estimators are consistent only if the
assumptions about the error term are valid.
 In cases of heteroscedasticity and non-normal error term, the ML
estimators are inconsistent.
 Alternative methods need to be devised in such situations as compute
robust standard errors,
 In practice, censored regression models may be preferable to the
truncated regression models because in the former we include all the
observations in the sample, whereas in the latter we only include
observations in the truncated sample.

Modeling count data: the Poisson and negative binomial
regression models
 The regressand is of the count type in many a phenomena, such as the
number of:
visits to a zoo in a given year,
patents received by a firm in a year,
visits to a dentist in a year,
speeding tickets received in a year,
cars passing through a toll booth in a span of, say, 5 minutes, and
so on. The
 underlying variable in each case is discrete, taking only a finite non-
negative number of values.
 Tests are conducted as in Logistic regression.
 When the mean and variance are not equal, the Poisson Distribution
will be replaced with Negative Binomial Distribution
 A unique feature of all these examples is that they take a finite number
of non-negative integer, or count, values.
 Regression models based on poisson probability distribution are
known as poisson regression models (PRM).
 An alternative to PRM is negative binomial regression model (NBRM),
which is based on the negative binomial probability distribution and is
used to remedy some of the deficiencies of the PRM.
 This model is nonlinear in parameters, necessitating nonlinear
regression estimation.
 This can be accomplished by the method of maximum likelihood (ML).
 The Poisson regression model which is often used to model count data,
based on the Poisson probability distribution(PPD.
 A unique feature of the PPD is that the mean of a poisson variable is the
same as its variance.This is also a restrictive feature of PPD.

 We cannot use the normal probability distribution to model count data.
 The Poisson probability distribution (PPD) is often used to model count
data, especially to model rare or infrequent count data.

2.11 Tabulation of patent raw data.
Tabulation ofP90
Sample: 1 181
Number of categories: 5
Cumulative Cumulative
# Patents Count Percent Count Percent
[0,200) 160 88.40 160 88.40
[200,400) 10 5.52 170 93.92
[400,600) 6 3.31 176 97.24
[600,800) 3 1.66 179 98.90
[800,1000) 2 1.10 181 100.00
Total 181 100.00 181 100.00

Table 2.12 Poisson model of patent data (ML estimation).
Dependent Variable: P90: Method: ML/QML - Poisson Count (Quadratic hill
climbing)Sample: 1 181, Included observations: 181 and Convergence achieved
after 6 iterations Covariance matrix computed using second derivatives
variable Coefficient Std. Error Z-Statistic Prob. Values
Constant -0.7458 0.0621 -12.003 0.0000
LR90 0.8651 0.0080 107.23 0.0000
AEROSP -0.7965 0.0679 -11.721 0.0000
CHEMIST 0.7747 0.0231 33.500 0.0000
COMPUTER 0.4688 0.0239 19.586 0.0000
MACHINES 0.6463 0.0380 16.994 0.0000
VEHICLES -1.1556 0.0391 -38.432 0.0000
JAPAN -0.0038 0.0268 -0.1449 0.8848
US -0.4189 0.0230 -18.140 0.0000
Log likelihood is -5081, Restr. log likelihood is -15822 and LR statistic is 21482
with its p.v. of 0.0000,
.

Limitation of the Poisson regression model
 The Poisson regression results for the patent and R&D given in Table 12.4
should not be accepted at face value.
 The standard errors of the estimated coefficients given in that table are
valid only if the assumption of the Poisson distribution underlying the
estimated model is correct.
 If there is over-dispersion, the PRM estimates, although consistent are
inefficient with standard errors that are downward biased.
 If this is the case, the estimated Z values are inflated, thus over-estimating
the statistical significance of the estimated coefficients.

 The main point to note is that if one uses the poisson regression
model(PRM), it is subject to over-dispersion test.
 If the test shows over-dispersion, we should at least correct the standard
errors by method of quasi-maximum likelihood estimation (QMLE) and
generalized linear model(GLM).
 If the assumption of equi-dispersion underlying the PRM cannot be
sustained, and
 Even if we correct the standard errors obtained by ML, it might be better
to search for alternative to PRM, which is the negative binomial regression
model, based on the negative binomial probability distribution.

 But these results may not be reliable because of the restrictive
assumption of the PPD that its mean and variance are the same.
 In most practical applications of PRM, the variance tends to be greater
than the mean. This is the case of over-dispersion.
 To correct for over-dispersion, we can use the methods of quasi
maximum likelihood estimation (QMLE) and generalized linear model
(GLM).
 Both these methods correct the standard errors of the PRM, which
was estimated by the method of maximum likelihood (ML).
 As a result of these corrections, it was found that several standard
errors in the PRM were severely underestimated, resulting in the
inflated statistical significance of the various regressors.

 In some cases, the regressors are found to be statistically insignificant,
in strong contrast with the original PRM estimates.
 Since our results showed over-dispersion, we used an alternative
model, the negative binomial regression model (NBRM).
 An advantage of NBRM model is that it allows for over-dispersion and
also provides a direct estimation of the extent of overestimation of the
variance.
 The NBRM results also show that the original PRM standard errors
were underestimated in several cases.

Chapter 2

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Chapter 2

Uploaded by

Copyright:

Available Formats

CHAPTER 2

QUALITATIVE EXPLANATORY AND REGRESSAND VARIABLE

8/1/2021 For MSc students by Urgaia R.(Ph.D.),OSU 1

 In regression analysis , in addition to quantitative variables, the

8/1/2021 For MSc students by Urgaia R.(Ph.D.),OSU 2

8/1/2021 For MSc students by Urgaia R.(Ph.D.),OSU 4

8/1/2021 For MSc students by Urgaia R.(Ph.D.),OSU 5

8/1/2021 For MSc students by Urgaia R.(Ph.D.),OSU 6

8/1/2021 For MSc students by Urgaia R.(Ph.D.),OSU 14

8/1/2021 For MSc students by Urgaia R.(Ph.D.),OSU 16

Dependent Variable: WAGE, Method: Least Squares, Sample: 1 1289

8/1/2021 For MSc students by Urgaia R.(Ph.D.),OSU 19

8/1/2021 For MSc students by Urgaia R.(Ph.D.),OSU 22

8/1/2021 For MSc students by Urgaia R.(Ph.D.),OSU 23

Cons -10.645 1.3718 -7.7600 0.0000

8/1/2021 For MSc students by Urgaia R.(Ph.D.),OSU 24

8/1/2021 For MSc students by Urgaia R.(Ph.D.),OSU 30

8/1/2021 For MSc students by Urgaia R.(Ph.D.),OSU 31

8/1/2021 For MSc students by Urgaia R.(Ph.D.),OSU 32

8/1/2021 For MSc students by Urgaia R.(Ph.D.),OSU 37

8/1/2021 For MSc students by Urgaia R.(Ph.D.),OSU 44

8/1/2021 For MSc students by Urgaia R.(Ph.D.),OSU 45

𝑙𝑛𝐿 = 𝑦𝑖 𝑙𝑛 𝐺(𝑋𝑖 𝛽) + (1 − 𝑦𝑖 )𝑙𝑛 1 − 𝐺(𝑋𝑖 𝛽)

This equation is highly non-linear, and we need to apply numerical optimization

8/1/2021 For MSc students by Urgaia R.(Ph.D.),OSU 46

𝑙𝑛𝐿 = 𝑦𝑖 𝑙𝑛 𝐺(𝑋𝑖 𝛽) + (1 − 𝑦𝑖 )𝑙𝑛 1 − 𝐺(𝑋𝑖 𝛽)

8/1/2021 For MSc students by Urgaia R.(Ph.D.),OSU 47

𝑙𝑛𝐿 = 𝑦𝑖 𝑙𝑛 𝑒 𝛼 +𝑋𝑖 𝛽 − 𝑦𝑖 ln⁡

The first order conditions are:

8/1/2021 For MSc students by Urgaia R.(Ph.D.),OSU 48

8/1/2021 For MSc students by Urgaia R.(Ph.D.),OSU 49

We realize that the sigmoid, or S-

Fig. A cumulative distribution function (CDF) can be both for logit

8/1/2021 For MSc students by Urgaia R.(Ph.D.),OSU 51

•If entire interval is above 1, conclude positive association

8/1/2021 For MSc students by Urgaia R.(Ph.D.),OSU 52

8/1/2021 For MSc students by Urgaia R.(Ph.D.),OSU 53

8/1/2021 For MSc students by Urgaia R.(Ph.D.),OSU 55

8/1/2021 For MSc students by Urgaia R.(Ph.D.),OSU 60

8/1/2021 For MSc students by Urgaia R.(Ph.D.),OSU 62

McFadden R-squared = 0.0297, LR statistic= 47.26 with Prob(LR statistic)=0.000

8/1/2021 For MSc students by Urgaia R.(Ph.D.),OSU 64

8/1/2021 For MSc students by Urgaia R.(Ph.D.),OSU 65

8/1/2021 For MSc students by Urgaia R.(Ph.D.),OSU 66

8/1/2021 For MSc students by Urgaia R.(Ph.D.),OSU 68

McFadden R-squared =0.0377, LR statistic= 59.96 with Prob(LR statistic)=0.000

8/1/2021 For MSc students by Urgaia R.(Ph.D.),OSU 69

8/1/2021 For MSc students by Urgaia R.(Ph.D.),OSU 70

8/1/2021 For MSc students by Urgaia R.(Ph.D.),OSU 71

HUSWAGE -105.509 15.629 -6.7507 0.0000

8/1/2021 For MSc students by Urgaia R.(Ph.D.),OSU 74

8/1/2021 For MSc students by Urgaia R.(Ph.D.),OSU 75

8/1/2021 For MSc students by Urgaia R.(Ph.D.),OSU 76

8/1/2021 For MSc students by Urgaia R.(Ph.D.),OSU 77

8/1/2021 For MSc students by Urgaia R.(Ph.D.),OSU 79

8/1/2021 For MSc students by Urgaia R.(Ph.D.),OSU 80

8/1/2021 For MSc students by Urgaia R.(Ph.D.),OSU 82

8/1/2021 For MSc students by Urgaia R.(Ph.D.),OSU 85

8/1/2021 For MSc students by Urgaia R.(Ph.D.),OSU 86

8/1/2021 For MSc students by Urgaia R.(Ph.D.),OSU 90

8/1/2021 For MSc students by Urgaia R.(Ph.D.),OSU 92

8/1/2021 For MSc students by Urgaia R.(Ph.D.),OSU 93

8/1/2021 For MSc students by Urgaia R.(Ph.D.),OSU 96