Econometric Method

Midterm 2 - Take Home Sectio

First Name: Minir
Last Name: Jafarov
No: 2172902
1. y = β0 + β1 x1 + β2 x 2 + β3 x3 + u - PR
E(u | x1, x2, x3) = 0
y = β0 + β1 x1 + β2 x2 + v - Researcher’s PR
v = β3 x3 + u
We know that when Researcher’s PRF is estimated, we get
ỹ = β˜0 + β˜1 x1 + β˜2 x2
Also, we know that
∑ ri1̂ yi
β˜1 =
∑ r 2i1̂
ri1̂ is the OLS residual collected from regression of x1 on x2 . We ca plug real PRF,
y = β0 + β1 x1 + β2 x2 + β3 x3 + u, into β˜1 function
∑ ri1̂ ⋅ (β0 + β1 x1 + β2 x2 + β3 x3 + u) ∑ ri1̂ ∑ ri1̂ xi1 ∑ ri1̂ xi2 ∑ ri1̂ xi3 ∑ ri1̂ ui
β˜1 = = β0 + β1 + β2 + β3 +
̂
∑ r i1
2 ̂
∑ r i1
2 ∑ r i1̂
2 ∑ r i1̂
2 ∑ r i1̂
2 ∑ r 2i1̂

From here, we understand tha

ri1̂ = 0, ri1̂ xi1 = r 2i1

̂ , and ri1̂ xi2 = 0.
∑ ∑ ∑ ∑
So, our β˜1 function becomes
∑ ri1̂ xi3 ∑ ri1̂ ui
β˜1 = β1 + β3 +
̂
∑ r 2i1 ∑ r 2i1̂
Now, let’s calculate expected value of β˜1 conditional on x1 , x 2 , and x3 . After that we can
check if it is biased or not.

( ̂ ) ( ∑ r 2i1
̂ )
∑ ri1̂ xi3 ∑ ri1̂ ui ∑ ri1̂ xi3 ∑ ri1̂ E(ui )
E( β˜1 ) = E(β1) + E β3 + E = β1 + β3 +
∑ r 2i1 ̂
∑ r 2i1 ̂
∑ r 2i1
Based on assumption, E(ui ) = 0. Thus,
∑ ri1̂ xi3
E( β˜1 ) = β1 + β3
̂
∑ r 2i1
The bias is: E( β˜1 ) − β1
8

 ∑ ri1̂ xi3 ∑ ri1̂ xi3

Bia s = β1 + β3 − β1 = β3
̂
∑ r 2i1 ̂
∑ r 2i1
2. a) log(wage) = β0 + β1educ + β2 IQ + β3ex per + β4 ex per 2 + β5tenure+
+β6tenure 2 + β7 m ar r ied + u
From the dataset, we have generated ex persq and tenuresq on STATA for our regression
model. The results of regression is as follows:

log(wage) = 5.00729 + 0.05514educ + 0.00543IQ + 0.01902ex per − 0.00018ex per 2+

+0.02261tenure − 0.00065tenure 2 + 0.1958m ar r ied
All variables, except for ex per , ex persq , and tenuresq , seem statistically signi cant at
5%
One additional year of schooling is going increase wage by 5.51%
One point increase in intelligence test result will increase wage by 0.54%
One more year in labour market will expand wage by 1.9%
According to this model, experience has diminishing marginal effect on wage since β3̂ > 0
and β4̂ < 0 . Experience rst increases wage, but after certain point wage starts to decrease.
Turning point of experience is

β3̂ 0.01902
ex per* = = = 52.83
2β4̂ 2 ⋅ (−0.00018)

One additional year with the same employer will increase wage by 2.26%
According to this model, tenure also has diminishing marginal effect on wage since β5̂ > 0
and β6̂ < 0 . Wages rst increase with tenure, but after certain point wage starts to decrease.
Turning point for tenure is
.

fi
fi

fi
 β5̂ 0.02261
tenure* = = = 17.39
2β6̂ 2 ⋅ (−0.00065)

Lastly, it’s clear that there is marriage premium. On average, married men earn 19.58% more
than single men, ceteris paribus. More speci cally, if we take a married man and a single man
with the same level of education, IQ, experience, and tenure, the married man will earn on
average 19.58% more than single man
b) We can test the model and see if it is jointly statistically signi cant
H0 : β1 = β2 = . . . = β7 = 0
H1 : at least one is different

By running STATA, we can see that F-value of the model is 34.16 and the p-value is almost
0. So, we reject H0 and the model is jointly statistically signi cant
c) For checking the return to an additional year of education, we need to create interaction
term. We generate m ar r ied−educ on STATA and regress the model. The results are:
.

fi
.

fi
fi
.

 Estimated return to education for single men is 3.53%. However, for married men, the return
to education is 0.03536 + 0.02227 = 0.05763 which means 5.76%.
As can be seen, in this model, more variables became statistically insigni cant. Also, the
coef cient of m ar r ied shows the difference in wages between single and married men who
have educ = 0 which is not realistic. Moreover, since there is high multicollinearity between
m ar r ied and m ar r ied−educ , the standard errors are high. So, it can be more useful to
look at the model with the mean education level.

The mean education level is 13.46 as can be seen. So, we will generate new variable which
would be m ar r ied−educ1 = m ar r ied * (educ − 13.46).

As it is obvious from the coef cient of m ar r ied, married men on average earn 18.8% more
wage than single men on mean education level. But since m ar r ied−educ1 is statistically
insigni cant, we can jointly test it.

It’s clear that jointly, they’re statistically signi cant

d) Now let’s estimate the regression models for single and married men separately
fi
fi
fi
fi

fi
.

For Married Men:

For Single Men:

Also, let us have larger model including interaction terms for all variables and estimate it.
Let’s test the equality using F-test:

We can reject the equality of the parameters at 5% level, even at 1%

We can use another method of Chow test
[SSRP − (SSR1 + SSR2 )] [n − 2(k + 1)]
F= ⋅ .
SSR1 + SSR2 k +1

SSR(restricted) = SSRP = 135.051341

SSR(married) = SSR1 = 117.30165
SSR(single) = SSR2 = 13.3467311

As we know all necessary values, we can calculate Chow’s F-value

This gives us similar result and we reject H0 which states that the regression function is the
same between married and single men
Also, we can look at the version of Chow Test where we omit married dummy and test only
interactions.
.

 It’s clear that in this case, we fail to reject H0, so it is better to include just marital status
instead of interactions
e) From the graph below, it’s clear that there is possibility of heteroskedasticity because
points are not gathered around 0. So, it’s useful to test if there is heteroskedacticity or not.
2
1
Residuals
0-1
-2

6.2 6.4 6.6 6.8 7 7.2

Fitted values

Let’s rst run Breusch-Pagan test.

As can be seen, at 5% level, we reject the H0 . Thus, we can assume that there is
heteroskedacticity in this model
f) We can check heteroskedasticity also by using White test:

As it’s clear, here, we have the same result as in Breusch-Pagan test. The only difference is
that, in Breusch-Pagan test, we failed to reject H0 at 1% level, while in White test, we reject
null hypothesis of homoskedasticity at any level
g) We can also check the model to see if there’s any kind of misspeci cation. We can use
Ramsey’s RESET test.
fi
.

fi
 As can be seen, the p-value is so high. Thus, we fail to reject H0 which states that there is no
misspeci cation.
To overall, our model suffers from heterosckedasticity, however, it’s correctly speci ed
3. Since there is heterosckedasticity, using OLS standard errors and test statistics are
misleading. Instead of it, we can estimate Heteroskedasticity-robust standard errors, under
which t-statistics will be valid.

Except from it, we can use Bootstrap standard errors under heteroskedasticity.

As can be seen, Heteroskedasticity-robust standard errors and Bootstrap standard errors are
almost the same. Only Heteroskedasticity-robust standard errors are a little bit larger. But it’s
not the problem. Both of these methods can be used to overcome heteroskedasticity problem
in the model
fi
.

fi
.