Scribd
Upload
95 views

Advanced Econometrics - Assignment 2

This document discusses a regression model of wage on several variables and the testing and correction of heteroskedasticity in the model. (1) The initial OLS model showed that gender significantly predicted the variance of the error term, violating an assumption of OLS. (2) Tests confirmed the presence of heteroskedasticity. (3) Using feasible generalized least squares (FGLS) to model the varying error variance, coefficients declined but significance was unchanged, correcting for heteroskedasticity.

Uploaded by

lolDevR
Copyright
© © All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as PDF, TXT or read online on Scribd
95 views

Advanced Econometrics - Assignment 2

This document discusses a regression model of wage on several variables and the testing and correction of heteroskedasticity in the model. (1) The initial OLS model showed that gender significantly predicted the variance of the error term, violating an assumption of OLS. (2) Tests confirmed the presence of heteroskedasticity. (3) Using feasible generalized least squares (FGLS) to model the varying error variance, coefficients declined but significance was unchanged, correcting for heteroskedasticity.

Uploaded by

lolDevR
Copyright
© © All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as PDF, TXT or read online on Scribd
You are on page 1/ 2

Advanced Econometrics - Assignment 2

Devvrat Raghav

The log file for this assignment has been emailed.

(a.) The initial model is:

wage = b0 + b1 educ + b2 exper + b3 expersq + b4 f emale + b5 married + u

For the error variance to rely only upon the individual’s gender, our model of V ar(u) must assume
that the variance of u, conditioned on all the regressors, is the same as the variance of u conditioned
on the dummy for gender, i.e.

V ar(u | educ, exper, expersq, f emale, married) = V ar(u | f emale) = β0 + β1 f emale

According to the above model, the variance of u for males is β0 , while for females it is β0 + β1 .
However, we can’t observe the u’s directly, and so we must estimate them. Since the best unbiased
estimator of V ar(u) is û2 , we have:

û2i = β0 + β1 f emale + ei

(b.) Regressing the above model for error variance on Stata, we get:

û2i = 11.898 − 6.379f emale + ei

Hence, our model shows that the estimated error variance is smaller for females than for males. This
result is significant for a 0.1% level of confidence, as the p-value for the t-test of β1 is = 0.001.

(c.) The error term in the initial regression model is heteroskedastic. For starters, we found that the
individual’s gender is a significant predictor of the error term’s variance. This invalidates MLR.5 for
this model, as the error variance is not independent of the regressors.

Similarly, results from the Breusch-Pagan Test and White’s Test for Heteroskedasticity compel us to
reject the null hypothesis of homoskedasticity (both p-values are equal to 0).

(d.) The robust standard errors are not too different from the results obtained in part b). However, it
is notable that the standard errors for all beta coefficients, except b0 and b1 , increased when calculated

1
via heteroskedasticity-robust methods. Even so, the statistical significance (or lack thereof) of all the
regressors remains unchanged at conventional levels of significance (10%, 5% and 1%).

(e.) Since the error term’s variance now relies on two regressors and has an unknown functional
form, we must use Feasible GLS to try and model V ar(u|x). Hence, the model for error variance
becomes:
V ar(u) = σ 2 exp(α0 + α1 educ + α2 f emale)

By regressing ln(û2 ) on educ and f emale, we obtain residuals that are exponentiated and inverted.
We use the square root of these transformed residuals (wt) as the weights to estimate the regression
model for wage. Consequently, we find that all beta coefficients fall after wt becomes a part of the
regression equation. This decline is especially sharp for educ and exper. This difference in coefficients
(compared to OLS) could be explained by the fact that FGLS estimates are not unbiased, but simply
consistent and asymptotically more efficient than OLS.

Notably, no tangible change in the significance of the regressors was observed. This is a result of the
standard errors declining in value once heteroskedasticity is corrected-for, which yields statistically
significant t-statistics. The only real exception is married, since it becomes even more insignificant.
However, since it coefficient (b5 ) isn’t very different in magnitude from 0, we don’t place much
emphasis on this result.

You might also like