Recap

I We have studied the multiple regression model and learnt that when:
1. model is linear in parameters: y = Xβ + u
2. conditional mean of errors is zero: E (u | X) = 0
3. columns of X are linearly independent
⇒ then the OLS estimator βb is an unbiased estimator of β

2 / 27
Recap
I We have studied the multiple regression model and learnt that when:
1. model is linear in parameters: y = Xβ + u
2. conditional mean of errors is zero: E (u | X) = 0
3. columns of X are linearly independent
⇒ then the OLS estimator βb is an unbiased estimator of β
I if in addition,
4. sample is random and errors are homoskedastic: Var (u | X) = σ 2 In ,
b = σ 2 (X0 X)−1
⇒ then βb is the BLUE and Var (β)

2 / 27
Recap
I We have studied the multiple regression model and learnt that when:
1. model is linear in parameters: y = Xβ + u
2. conditional mean of errors is zero: E (u | X) = 0
3. columns of X are linearly independent
⇒ then the OLS estimator βb is an unbiased estimator of β
I if in addition,
4. sample is random and errors are homoskedastic: Var (u | X) = σ 2 In ,
b = σ 2 (X0 X)−1
⇒ then βb is the BLUE and Var (β)
I If, in addition to the above,
5. errors are normally distributed,
⇒ then conditional on X, βb is normally distributed, and we can use the
usual t and F tests to make inferences based on the OLS estimator

2 / 27
Lecture Outline
I Heteroskedasticity:
1. Definition of heteroskedasticity and its consequences for OLS
(textbook reference 8-1)
2. Testing for heteroskedasticity (textbook reference 8-3)
2.1 Breusch-Pagan test
2.2 White test
3. Heteroskedasticity robust standard errors (a simplified version
of 8-2)
4. Weighted least squares when heteroskedasticity is known up to
a multiplicative constant (textbook reference 8-4a)
I We will not cover heteroskedasticity robust LM tests (the last part
of section -.2), Feasible GLS and the consequences of wrong
specification of the variance function (section 8.4b, 8.4c) and the
linear probability model (section 8-5).

3 / 27
Heteroskedasticity
I Sometimes there is a good reason to doubt the assumption of equal
variance for all errors. Here are some examples:
I In the study of food consumption, income is an important
explanatory variable. It is unreasonable to assume that the variance
of food consumption is the same for poor and rich people
I In many cases we do not have individual data (for confidentiality
reasons), but we get information on averages over groups of
individuals.
I For example, we can get incidences of crime per 1000 people,
employment rate and income per capita in each district. These are
averages, but different districts have different populations, so there
is a good reason to believe that variances of these averages depend
inversely on the population of each district
I In finance, some unpredicted news increase the volatility of the
market (i.e. the variance of the market return) and this can last for
several days (a large part of financial econometrics ETC3460 is
about modelling this phenomenon)
4 / 27
I A 3D graphical representation of heteroskedasticity:

I In this example, the variance of u is getting larger as x increases

5 / 27
Consequences of heteroskedasticity (HTSK) for OLS

I HTSK does not affect the first 3 assumptions on the Recap slide,
therefore the OLS estimator will remain unbiased
I HTSK violates assumption 4, therefore the OLS estimator will no
longer be BLUE and Var (β) b 6= σ 2 (X0 X)−1 . This means that the
default standard errors reported by the statistical package for the
OLS estimator will be incorrect
I Even if errors are Normally distributed, the t and F tests based on
the default OLS standard errors will be unreliable
I Fortunately, if we detect HTSK, we have ways to conduct reliable
inference based on the OLS estimator, or even obtain a more
efficient estimator than the OLS estimator

6 / 27
Detecting HTSK
I As always, step 1: think about the problem!
I If we only have one x, the scatter plot can give us a clue:

I However, we hardly ever have only one x :-(

7 / 27
Testing for HTSK
I Since E (ui | xi1 , . . . , xik ) = 0,
Var (ui | xi1 , . . . , xik ) = E (ui2 | xi1 , . . . , xik )
I If we suspect that variance can change with some subset of the
independent variables, or even some exogenous variables that do not
affect the mean, but can affect the variance, then, if we had ui , we
could square it and estimate the conditional expectation function of
ui2 . But ui is unknown :-(
I Australian econometricians Trevor Breusch and Adrian Pagan, and
the American econometrician Hal White showed that we can use the
OLS residuals instead, and in large samples this will give us reliable
results :-)

8 / 27
Breusch-Pagan test
yi = β0 + β1 xi1 + β2 xi2 + · · · + βk xik + ui for i = 1, . . . , n
H0 : E (ui2 | xi1 , xi2 , . . . , xik ) = σ 2 for i = 1, . . . , n
H1 : E (ui2 | xi1 , xi2 , . . . , xik ) = δ0 + δ1 zi1 + δ2 zi2 + · · · + δq ziq
where zi1 , . . . , ziq are a subset of xi1 , xi2 , . . . , xik . In fact the z variables
can include some variables that do not appear in the conditional mean,
but may affect the variance (note the difference with the book)
1. Estimate the model by OLS as usual. Obtain ûi for i = 1, . . . , n and
square them.
2. Regress ûi2 on a constant, zi1 , . . . , ziq . Denote the R 2 of this
auxiliary regression by Rû22 .
3. Under H0 , the statistic n × Rû22 has a χ2 distribution with q degrees
of freedom in large samples. This statistic is called the Lagrange
multiplier (LM) statistic for HTSK.
4. Given the desired level of significance, we obtain the cv of the test
from the χ2 table, and reject H0 if the value of the test statistic is
larger than the cv.
9 / 27
I Under H0 the F test for the overall significance of the second
regression has an Fq,n−q−1 distribution, and it can also be used to
test for HTSK.
I Example: Net financial wealth in $1,000s, predicted by age and
current income in $1,000s

10 / 27
I Can save residuals, and run the OLS of û 2 on a constant and all the
explanatory variables, or we can use Eviews

11 / 27
I Choosing Breusch-Pagan test in Eviews produces:

I When learning, it is better to do the LM steps yourself rather than

to use Eviews options in order to make sure you understand the
mechanics of the test

12 / 27
White Test for HTSK

I White’s test of course tests the same null hypothesis

H0 : E (ui2 | xi1 , xi2 , . . . , xik ) = σ 2 for i = 1, . . . , n

but its alternative is that the variance is a smooth unknown

function of xi1 , xi2 , . . . , xik .
I Hal White showed that a regression of û 2 on a constant, x1 to xk ,
x12 to xk2 and all pairwise cross products of x1 to xk , has the power
to detect this general form of heteroskedasticity in large samples.
I Similar to the B-P test, White’s test statistic is n × Rû22 , which
under the null, has a χ2 distribution with degrees of freedom as the
number of explanatory variables in the auxiliary regression. We can
also use the F-test of the overall significance of the auxiliary
regression as well.

13 / 27
A Special Case of the White Test for HTSK

I A concern with White test is that the auxiliary regression will have
k + k(k + 1)/2 regressors, which is a very large number of
restrictions
I Recall that the fitted values from OLS, ŷ are a function of all the xs
I Thus, ŷ 2 will be a function of the squares and crossproducts of all
the x’s and ŷ and ŷ 2 can proxy for all of the the x’s their squares
and cross products.
I A special form of the White test would be to regress the residuals
squared on ŷ and ŷ 2 and use the R 2 of this regression to form an F
or LM statistic
I Note only testing for 2 restrictions now, no matter how many
independent variables we have

14 / 27
I In the financial wealth example, choosing the White test in Eviews
produces:

15 / 27
I Eviews does not have a built in command for the alternate form of
the White test, and we need to save the OLS residuals and OLS
predictions and then run the auxiliary regression:

I The value of the n × R 2 statistic is 2017 × 0.0078 = 15.73 which is

larger than 5.99, the 5% critical value of the χ2 distribution with 2
degrees of freedom. All tests lead to the same conclusion: There is
significant evidence of HTSK, so, although OLS is still unbiased, it
is no longer the best and OLS standard errors are unreliable

16 / 27
Solution 1: Robust Standard Errors
I Since OLS estimator is still unbiased, we may be happy to live with
the OLS even if it is not BLUE. But the real practical problem is
that t and F statistics based on OLS standard errors are unusable
I Recall the derivation of Var (βb | X):

Var (βb | X) = (X0 X)−1 [X0 Var (u | X)X] (X0 X)−1

I With homoskedasticity,

Var (u | X) = σ 2 In ⇒ Var (βb | X) = σ 2 (X0 X)−1

I With HTSK
 2 
σ1 0 ··· 0
0 σ22 ··· 0
Var (u | X) =  .
 
.. .. .. 
 .. . . .
0 0 ··· σn2

17 / 27
I Therefore, with HTSK
  2  
σ1 0 ··· 0
 0 σ22 ··· 0 
Var (βb | X) = (X0 X)−1 X0  .
  0 −1
X  (X X)
 
.. .. ..
  ..

. . . 
0 0 ··· σn2

I Amazingly, White proved that:

  2  
û1 0 ··· 0
 0 û22 ··· 0 
d (βb | X) = (X0 X)−1 
Var X0  ..

.. .. ..
  0 −1
 X  (X X)
 . . . . 
0 0 ··· ûn2

is a reliable estimator for Var (βb | X) in large samples

I The square root of diagonal elements of this matrix are called White
standard errors or robust standard errors, which most statistical
packages compute. These are reliable for inference.

18 / 27
I Back to the example. The option of robust standard errors is under
the Options tab of the equation window:

19 / 27
I With this option, we get the following results:

I Compare with the original regression results:

20 / 27
Solution 2: Transform the Model

a. Logarithmic transformation of y may do the trick: If the population

model has log(y ) as the dependent variable but we have used y ,
this kind of mis-specification can show up as heteroskedastic errors.
So, if log-transformation is admissible (i.e. if y is positive), moving
to a log model may solve the problem, and the OLS estimator on
the log-transformed model will then be BLUE and standard errors
will be useful. Of course when we consider transforming y , we
should think if a log-level or a log-log model makes better sense
b. Weighted least squares: When there is good reason to believe that
variance of each error is proportional to a known function of a single
independent variable, then we can transform the model in a way to
eliminate HTSK and then use OLS on the transformed model. This
estimator is the weighted least squares (WLS) estimator, which we
derive on the next slide.

21 / 27
Weighted Least Squares
I Suppose the model

yi = β0 + β1 xi1 + β2 xi2 + · · · + βk xik + ui for i = 1, . . . , n (1)

satisfies the assumptions needed for unbiasedness of OLS, and we

have
Var (ui | xi1 , xi2 , . . . , xik ) = σ 2 hi
where hi is a known function of one of x’s, or a function of a
variable z as long as E (ui | xi1 , xi2 , . . . , xik , zi ) = 0. For example
2
hi = xi1 , or hi = xi1 , or hi = 1/zi .
I Multiplying both sides of equation (1) by wi = √1 eliminates
hi
HTSK because:

22 / 27
I The transformed (or “weighted”) model:

(wi yi ) = β0 wi + β1 (wi xi1 ) + β2 (wi xi2 ) + · · · + βk (wi xik ) + (wi ui ) (2)

satisfies all assumptions of the Gauss-Markov theorem, so the OLS

estimator of its parameters is BLUE.
I More importantly, equation (2) has the same parameters as
equation (1). So, OLS on the weighted model will produce BLUE of
β0 to βk and we can test any hypotheses on these parameters based
on the weighted model.
I Note that the transformed model does not have a constant term,
and β0 is the coefficient of wi in the transformed model
I This estimator is called the weighted least squares (WLS) estimator
of β

23 / 27
I In the financial wealth example, the auxiliary regression suggests
that the variance changes with income. Since income is positive for
all observations (why is this important?), we hypothesise that
Var (ui | inci , agei ) = σ 2 inci
I We create wi = √ 1 [Eviews command: series
inci
w=1/@sqrt(inc)] and we run the weighted regression

I The standard errors are now reliable for inference and for forming
confidence intervals

24 / 27
I Eviews also has a built in WLS command under the option tab of
the equation window. We need to enter the name of the weight
series.

25 / 27
I The only advantage of the Eviews built in command is that it
produces a set of statistics for the original model

I Note that while heteroskedasticity is more common in

cross-sectional data, it can also arise in time series data and
adjustments to standard errors apply.
26 / 27
Summary

I The assumption of equal conditional variances for each observation

may not be appropriate
I But OLS will still be unbiased even if the errors are heteroskedastic,
however the usual OLS standard errors will not be correct
I We learnt how to test for heteroskedasticity
I If HTSK is found, we can still use OLS, but calculate standard
errors that are robust to HTSK, and use those for inference
I If we have a reasonable idea that the HTSK is proportional to a
single variable, we can use WLS, which will provide the best linear
unbiased estimator for the parameters and a set of standard errors
that can be used for inference

27 / 27