Econometrics

University of Milan-Bicocca

Course lecturer:
Maryam Ahmadi
[email protected]

1
 Problem 20 & Answer.
1- Suppose in the following model, 𝑦=𝛽0 + 𝛽1 𝑥+𝑢, the variance of 𝑢 changes given different values of x
(for different individuals), so Var(𝑢𝑖 | 𝑥𝑖 )=𝜎𝑖2 .

a) Does this affect unbiasedness of OLS? Explain.

No, it doesn’t. This violated homoskedasticity (assumption A3), while for unbiasedness we need assumption
A1 and A2 to hold. Therefore, here, OLS is still unbiassed.

b) In the condition explained above, does the estimation of error variance

remain unbiased?

No. because here A3 does not hold and for unbiassed error variance we need all assumptions A1 to A4.

2
2- Suppose we have a model where we regress the firms export on age and size of the firms and we know that
this is the true model, we call it model 1: export= 𝛽0 + 𝛽1 𝑎𝑔𝑒+ 𝛽2 𝑠𝑖𝑧𝑒+u.

a) Suppose in model 1, size and age are highly correlated. Does this make the OLS biased? Explain your answer.
No. There is high multicollinearity among explanatory variables and this does not violate any of the linear
multiple regression assumptions, therefore, OLS remains unbiased.

b) In this case how will high correlation between size and age affect the statistical significance of size?
It affects the statistical significance of the effect of size through the t-statistic. Because multicollinearity
causes the variance of 𝛽መ2 to rise and therefore, the related t statistic tends to be small. In this case, it will be
very difficult to reject the null hypothesis and find a statistically significant effect of size.

3
c) Suppose we do not have data for the firms size, so we estimate a model without variable size and we call it
model 2: export= 𝛽0 + 𝛽1 𝑎𝑔𝑒+u. If we omit the size and estimate model2, what kind of problem can happen?
Omitting size causes the omitted variable bias. If the true model is: export= 𝛽0 + 𝛽1 𝑎𝑔𝑒+ 𝛽2 𝑠𝑖𝑧𝑒+u, and we drop size
and estimate: model 2, the estimated coefficient of size will be biased. This is called omitted variable bias. That is
𝐸 𝛽መ1 ≠ 𝛽1 . Given the omitted variable bias
0
෢ = 𝛽 + σ𝑛𝑖=1 𝑥1 ′ 𝑥1
E(𝛽) −1 σ𝑛 ′
𝑖=1 𝑥1 𝑥2 𝛽2 + σ𝑛𝑖=1 𝑥1 ′ 𝑥1 −1 σ𝑛𝑖=1 𝑥1 ′ 𝑢

We expect a positive effect from firm size to export (𝛽2>0) and a positive effect from firm age to firm size
(cov(𝑥1 , 𝑥2 )>0). Therefore, we expect a positive or upward bias in estimating the effect of age (𝛽1), or the effect will be
overestimated.

d) How do you expect the 𝑅2 of model 1 to be different from the 𝑅2 of model 2? Give a short answer.
R2 will never decrease if a variable is added. → We expect that 𝑅2 of model 1 to be larger 𝑅2 of model 2.

e) If we scale the export data from million dollars to billion dollars, how does it affect the 𝑅2 of the estimation? Give
a short answer.
The 𝑅2 will not change
4
3- We use a cross sectional dataset that includes information for 1,534 firms. The goal is to explain the
participation rate of firms’ workers in a specific pension plan. In this dataset:
prate: the percentage of workers that participate in the pension plan
mrate: the plan match rate that measures generosity of the pension plan (contribution by firm/contribution by worker).
age: age of the pension plan
totemp: total number of firm employees
In first step, we regress prate on mrate:
a) Compute the slope coefficient of the
estimation.

෡𝑗 −𝑞
𝛽
𝑡= ෡𝑗 , when H0: β𝑗 = 0, then q=0 and
se 𝛽

෡𝑗
𝛽
𝑡= ෡𝑗 → 𝛽መ𝑗 = t ∗ se 𝛽መ𝑗
se 𝛽

𝛽መ𝑗 = 0.5270107*11.12 = 5.860

5
b)Compute the 𝑅2 of the estimation.
2
σ 𝑦ො𝑖 − 𝑦ത 32001.7271
𝑅 = = = 0.07470
σ 𝑦𝑖 − 𝑦ത 428385.539

c) Compute the estimated error variance (𝜎̂2) of the model.

2 ෝ𝑖 2
෌ 𝑢 396383.812
𝜎ො = = = 258.73617
𝑁−𝑘−1 1534−1−1

d) Does this equation allow us to to draw “ceteris paribus” conclusion about how mrate affects prate?
No. Excluding totemp and age that are relevant variables, makes the error term correlated with both mrate and
prate. Therefore E(u|x)≠0 and zero conditional mean independence assumption is violated. The estimator will
be biased and the ceteribus paribus condition does not hold.

e) Find 95% confidence interval for the estimated coefficient of 𝑚𝑟𝑎𝑡𝑒.

𝐶. 𝐼. = 𝛽መ𝑗 ± 𝐶 ⋅ 𝑠e 𝛽መ𝑗 = 5.860358984±1.96*0.5270107=[4.8274,6.8933]

6
In second step, we regress prate on mrate, totemp and the squared term of totemp (totempsq).

f) Interpret the coefficient on mrate. Be sure to state the

null and alternative hypotheses, test statistic, critical
value, rejection rule, conclusion and interpretation.

𝐻0 : 𝛽𝑚𝑟𝑎𝑡𝑒 = 0 𝑎𝑛𝑑 𝐻1 : 𝛽𝑚𝑟𝑎𝑡𝑒 ≠0

t=11.00 and is larger than the 5% critical value with
df=1530 that is 1.96 (11>1.96). Therefore the effect of
mrate is significant at 5% level. This means that if mrate
increases by 1 unit, prate is expected to increase by 5.77
percentage point, keeping everything else constant.

g) What is the marginal effect of totemp on prate? (Express your answer in terms of β’s)
𝜕𝐸 𝑦𝑖 ȁ𝑥i
= 𝛽2 + 2. 𝛽3 .𝑡𝑜𝑡𝑒𝑚𝑝i
𝜕𝑡𝑜𝑡𝑒𝑚𝑝i

(h) Are the signs of 𝛽2 and 𝛽3 as expected? explain? Yes. Total number of employees has a negative but increasing
effect on prate. As the number of employees increases, its negative effect on prate tends to zero and after a
certain point will be positive. The higher the number of employees the higher percentage of worker that
participate in the pension plan.
7
In third step, we include age and the interaction term between age and mrate (mratge) to the model.

i) Test the hypothesis that there is no partial effect from age on prate. Be sure to state the null and alternative
hypotheses, test statistic, critical value, rejection rule and conclusion. (Be careful which coefficients you have to
consider for this test).

𝐻0 : 𝛽𝑎𝑔𝑒 = 𝛽𝑚𝑟𝑎𝑡𝑔𝑒 = 0 and 𝐻1 : 𝐻0 is not true.

(0.1132−0.0844)/2
𝐹(2,1528) = = 24.8118 > 3.00 → the null is rejected at 5% and we conclude that there is
(1−0.1132)/1528
partial effect from age to prate.
8
Good Luck!