VIOLATION OF CLRM ASSUMPTIONS:

HETEROSKEDASTICITY

Chapter 9
 HETEROSKEDASTICTIY: NATURE
 Usual OLS assumption is that variance of the conditional distribution of ,
given , is
 If instead, there is in the conditional variance of on , the error term is said
to be heteroskedastic. Symbolically:
 Homoskedasticity Population variance of residual terms, , remains
constant for all observations same dispersion
 Heteroskedasticity Population variance of the residual terms, , not
necessarily same for all  different dispersion
 Consider following 2-variable regression model:

Persona
l
savings Personal
disposable
income
 …HETEROSKEDASTICITY: NATURE

As PDI Although the

increases, the average level of
mean/average savings increases
level of savings as the PDI
also increases, increases, the
but the variance variance
of savings of savings does not
around its mean remain the same at
value remains all levels of PDI.
the same
at all levels of
PDI.
Therefore, in our regression of savings on income, error variances
(i.e., variance of ) associated with high-income families are
expected to be greater than those associated with low-income
families. Symbolically, we express heteroskedasticity as:
CONSEQUENCES OF HETEROSCEDASTICITY

1. OLS estimators are still linear.

2. They are still unbiased.
3. But they no longer have minimum variance; that is, they are no
longer efficient. OLS estimators are no longer BLUE in small as well as in
large samples (i.e., asymptotically).
4. The usual formulas to estimate the variances of OLS estimators are
generally biased. The bias can be positive (upward bias) or negative
(downward bias).

or

5. As a result, the usual confidence intervals and hypothesis tests based

on t and F distributions are unreliable.
 DETECTION OF HETEROSCEDASTICITY: GRAPHICAL EXAMINATION OF RESIDUALS
 Plot the squared residuals (residual plot) against X and check whether there is any
pattern

What about
linear
multivariable random heteroskedasti heteroskedasti
regression? c c
Plot squared
residuals
against each of
quadratic
or plot squared heteroskedasti
residuals c
against quadratic
estimated . heteroskedastic
 …DETECTION OF HETEROSCEDASTICITY: GRAPHICAL EXAMINATION OF RESIDUALS

Hypothetical patterns of estimated squared

residuals.
In this figure, are plotted against . The
idea being to find out whether the
estimated mean value of Y is
systematically related to the squared
residual. In part (a) we see that
there is no systematic pattern between
the two variables, suggesting that
perhaps no heteroscedasticity is
present in the data. Part (b), however,
exhibits definite patterns. Part (c)
suggests a linear relationship, whereas
Part (d) and (e) indicate a quadratic
relationship between and . Using such
informal knowledge, one may transform
6
the data in such a manner that the
transformed data do not exhibit
 …DETECTION OF HETEROSCEDASTICITY: PARK TEST
 If there is heteroscedasticity, the heteroscedastic variance may be systematically
related to one or more explanatory variables. To see if this is the case, we can regress
on one or more of the X variables.
 Park Test

1. Run the original regression: (1)

2. Obtain the residuals , square them, and take their logs
3. Run the regression: (2)
using the explanatory variable in the original model.
4. If there is more than one explanatory variable, run the regression against log of each
X variables. Alternatively, run the regression against, the estimated Y.
5. Test the null hypothesis that = 0; that is, there is no heteroscedasticity. If a
statistically significant relationship exists between and , the null hypothesis of no
heteroscedasticity can be rejected.
 Before we accept the results of the Park test, we should note some of the problems

associated with the test: The error term in regression (2), , may itself be
heteroscedastic.
DETECTION OF HETEROSCEDASTICITY: GLEJSER TEST
 The Glejser test allows us to explore the nature of heteroskedasticity a little more
closely than the Park Test. This test relax the assumption of Park test that estimated
error variance is proportional to variables only. Rather it investigates whether some
other functional form may be appropriate, for example,

(4)
(3)

(5)

 You may fit several such functions, varying the choice of . The null hypothesis in each
case is that there is no heteroscedasticity; that is, = 0. If this hypothesis is rejected,
there is probably evidence of heteroscedasticity.
 Glejser has found that for large samples the preceding models give generally
satisfactory results in detecting heteroscedasticity.
 As a practical matter, therefore, the Glejser technique may be used for large samples
and may be used in the small samples strictly as a qualitative device to learn
something about heteroscedasticity.
 DETECTION OF HETEROSCEDASTICITY: WHITE’S TEST
 White’s General Heteroscedasticity Test
 Suppose we have the regression model: (6)
 White’s test proceeds as follows:

1. Estimate above regression model by OLS, obtain the residuals, .

2. When the model contains independent variables, the White test is based on an
estimation of the following auxiliary regression:
(7)
3. Obtain the value from the auxiliary regression (7). Under the null hypothesis that
there is no heteroscedasticity (i.e., all the slope coefficients are zero), White has
shown that the value obtained from regression (7) times the sample size (), follows
the distribution with d.f. equal to the number of explanatory variables in regression
(7) (excluding the intercept term): . We can also use an F test of this hypothesis;
both tests have asymptotic justification.
4. If the value obtained from estimated equation exceeds the critical value at the
chosen level of significance, we can reject the null hypothesis of no
heteroscedasticity.
9
HETEROSKEDASTICITY: REMEDIAL MEASURES
 As we have seen, heteroscedasticity does not destroy the
unbiasedness property of OLS estimators, but the estimators are no
longer efficient, not even in large samples.
 This lack of efficiency makes the conventional OLS hypothesis testing
procedure of dubious value. Therefore, if heteroscedasticity is
suspected or diagnosed, it is important to seek remedial measures.
 If heteroskedasticity is confirmed by the Park, Glejser, White tests or
any other tests how can we solve it?
 If detected there are some ways in which we can transform the model,
so that in the transformed model there is no heteroskedasticity.
 The relevant question is what kind of transformation will work? This
answer depends on
(a) whether we know the true error variance
(b) whether the true error variance is unknown.
HETEROSKEDASTICITY: REMEDIAL MEASURES
 When is Known: The Method of Weighted Least Squares (WLS)

 To fix ideas, consider the following 2-variable regression

function:

(8)
 Where is heteroskedastic, so that, for all i. Now, to run WLS,
we need to transform the model by dividing it with known :

(9)
THE METHOD OF WEIGHTED LEAST SQUARES (WLS)

 Now, it is not hard to see whether the are homoscedastic:

Assuming all other CLRM
(since is known)
assumptions are satisfied, OLS
estimator of the transformed
model will be BLUE.
 Since error variance is now homoscedastic, application of OLS to (9)
will be efficient. However, two technical points need to be noted:
 Note that is no more a constant. Therefore, we need to run a

regression through origin.

 That is, we estimate two slope parameters, not as in (8), where

we need to estimate only one slope parameter. In fact, our model

in (9) is a multivariable regression model.
 The OLS estimators of and thus obtained are called weighted least
squares (WLS) estimators; each Y and X observation is weighted
(i.e., divided) by its own (heteroscedastic) standard deviation, .
DIFFERENCE BETWEEN OLS AND WLS UNDER HETEROSCEDASTICITY1

 From the definition of the disturbance term, we can see that the WLS
estimates are obtained by minimizing:
 (provided that is known)
 Whereas, in OLS we minimize:

 Thus, in WLS we weight (or, deflate) the importance of observations

with larger variance or RSS by larger variance. That is the larger the
variance, the larger the divisor will be.

13
 …HETEROSKEDASTICITY: REMEDIAL MEASURES
 When true is unknown
 The problem with the WLS procedure is that we do not know the disturbance
variance, , and usually need to make some assumptions about the nature of
the disturbance variance.
 Assumption 1: The error variance is proportional to : The square root
transformation:
Error variance proportional to
If after estimating the
X
usual OLS regression
we plot the residuals
against the
explanatory variable
X and observe a
(where k is a constant) pattern similar to that
shown in the figure,
this indicates that the
error variance
increases with the
increase in the value
…HETEROSKEDASTICITY: REMEDIAL MEASURES

 Transform the original model as follows:

(10)

 This transformation makes the disturbance term of (10)

homoscedastic or constant Therefore, one may proceed to apply OLS
 Since , to (10), regressing on and . Note an
important feature of the transformed
model: It has no intercept term.
 Therefore, Therefore, one will have to use the
regression-through-the-origin model to
 (using the assumption made)
estimate and . Having run (10), one can
get back to the original model simply by
multiplying (10) by .
 …HETEROSKEDASTICITY: REMEDIAL MEASURES

 Assumption 2: The Error Variance Proportional to :

 Suppose as a matter of “speculation,” graphical methods, or Park and
Glejser approaches, you believe that the variance of is proportional to the
square of the explanatory variable X (see the figure below), that is, . Assume
the original model you need to run is .
 Now, transform the model by dividing it by as follows:

Error variance proportional to
or,
Transformed
model
Where is transformed disturbance term, equal to . Now it is
easy to verify that
(using the assumption made)
 …HETEROSKEDASTICITY: REMEDIAL MEASURES

 Re-specification of the Model

 Instead of speculating about , sometimes a re-specification of the

regression model-choosing a different functional form-can reduce

heteroskedasticity.
 For example, instead of running the linear-in-variable (LIV)

regression, if we estimate the model in the log form, it often

reduces heteroskedasticity.
 Estimate

 Instead of
 Log transformation compresses the scales in which the variables

are measured.
 The number 90 is 10 times the number 9, but ln90 (= 4.4998) is

only about 2 times as large as ln9(= 2.1972).

HETEROSKEDASTICITY: APPLICATION OF TESTS

The table gives data on R&D

expenditure, and sales for 14
industry groups in the United States
($,million). Since the cross-sectional
data presented in this table are quite
heterogeneous, in a regression of
R&D on sales, heteroscedasticity is
likely. The OLS regression results are
as follows:

(11)

To see if our model suffers from

18
heteroskedasticity, we obtain and
from the model and plot them
…HETEROSKEDASTICITY: APPLICATION OF TESTS

It seems from this figure that there is a systematic pattern between

the residuals and squared residuals and sales, perhaps suggesting
that there is heteroscedasticity.
19
…HETEROSKEDASTICITY: APPLICATION OF TESTS

Park Test Glejser Test

(12) (13)

The Park test suggests that The Glejser test also suggests
there is a statistically significant that there is a systematic
positive relationship between relationship between the
squared residuals and sales. absolute values of the residuals
and sales, raising the possibility
that the regression (11) suffers
from heteroscedasticity

20
 …HETEROSKEDASTICITY: APPLICATION OF TESTS

Using the value and n = 14, we

obtain n = 6.090. Under the of no
White Test heteroscedasticity, this should follow a
(14) distribution (because there are two
regressors in (14). The p value of
obtaining a chi-square value of as
much as 6.090 or greater is about
0.0476. Since this is a low value, the
White test also suggests that there is
heteroscedasticity.

In sum, then, on the basis of the residual graphs and the Park, Glejser, and White
tests, it seems that our R&D regression (11) suffers from heteroscedasticity. Since
the true error variance is unknown, we cannot use the method of weighted least
squares to obtain heteroscedasticity-corrected standard errors and t values. 21