You are on page 1of 14

Outline

Introduction
Park test
Goldfeld-Quandt test
White test
Breusch Pagan test

Tests of Heteroscedasticity

Prof. Rizzi Laura

January 20, 2009

Prof. Rizzi Laura Tests of Heteroscedasticity


Outline
Introduction
Park test
Goldfeld-Quandt test
White test
Breusch Pagan test

Brief Overview on:


◮ Introduction
◮ Park test
◮ Goldfeld-Quandt test
◮ White test
◮ Breusch Pagan test

Prof. Rizzi Laura Tests of Heteroscedasticity


Outline
Introduction
Park test
Goldfeld-Quandt test
White test
Breusch Pagan test

Introduction

Econometricians do not use always the same test to verify the presence of heteroscedastic
disturbances, because heteroscedasticity may assume different structures and sometimes it is not
easy to understand this structure.
There is not a uniform approach in the choice of tests, but generally it is preferable to answer to
some questions before the application of whichever test:
◮ are there specification errors in the chosen regression model?
◮ are there possibilities of heteroscedastic errors in the phenomenon analized?
◮ considering the graphical distribution of residuals versus each regressor, is there evidence of
heteroscedasticity? It is interesting to analyse the graphical distribution of residuals versus
some regressors if they are thought to generate heteroscedasticity; if residuals appear related
(positively or negatively) with a regressor Z there may be heteroscedasticity.
◮ the assumption assumption of constant error variance can be checked throught the residual
plot. A residual plot is a scatterplot of the standardised residuals against the fitted values.

Prof. Rizzi Laura Tests of Heteroscedasticity


Outline
Introduction
Park test
Goldfeld-Quandt test
White test
Breusch Pagan test

Introduction

Residual plot - examples


The standardised residuals, si , are designed to overcome the problem of different variances of the
raw residuals. The problem is solved by dividing each of the raw residuals by an appropriate term.
Recall that the (standardised) residuals are the deviations of the observations away from the fitted
values. If Assumptions of constant error variance is satisfied we would expect the residuals to vary
randomly around zero and we would expect the spread of the residuals to be about the same
throughout the plot.

Residual plot for the relationship between ice


cream consumption and temperature, ice
cream price, average annual family income,
and the year.
The points in the plot seem to be fluctuating
randomly around zero in an un-patterned
fashion.
The plot does not suggest violations of the
assumption of constant variance of the
random errors.

Prof. Rizzi Laura Tests of Heteroscedasticity


Outline
Introduction
Park test
Goldfeld-Quandt test
White test
Breusch Pagan test

Introduction

If the residuals seem to increase or decrease in average magnitude with the fitted values, it is an
indication that the variance of the residuals is not constant.
If the points in the plot lie on a curve around zero, rather than fluctuating randomly, it is an
indication that linearity assumption is broken.
If a few points in the plot lie a long way from the rest of the points, they might be outliers, that is,
data points for which the model is not appropriate.

◮ Fig. a below shows a residual plot with


no systematic patter.
◮ In fig. b there is a clear curved pattern,
lineaqrity assumption may be broken.
◮ In fig. c the random variation of the
residuals increases as the fitted values
increase, then variance is not constant.
◮ Fig. d most of the residuals are
randomly scattered around 0, but one
observation has produced a residual
which is much larger than any of the
other residuals. The point may be an
outlier.

Prof. Rizzi Laura Tests of Heteroscedasticity


Outline
Introduction
Park test
Goldfeld-Quandt test
White test
Breusch Pagan test

Park test

This test procedure requires 3 steps:


◮ model OLS estimation to derive the OLS residuals, ei ;
◮ the derivation of the ln(ei2 which are considered as dependent variable in the regression
where the only regressor is the log of the r.v. considered proportionality factor;
◮ the estimation results of this model are used to verify the presence of heteroscedastic errors;
Then:
1 - Let consider the regression model: yi = β0 + β1 X1i + β2 X2i + ui ; the OLS estimation produces
the OLS residuals:
ei = yi − (b0 + b1 X1i + b2 X2i )

2 - We derive the dependent variable ln(ei2 ) for the regression:

2
ln(ei ) = α0 + α1 lnZi + vi

Where Z is a r.v. that may cause heteroscedasticity.


3 - We verify significance of the coefficient α1 using t test. If this coefficient is significant there is
heteroscedasticity explained by the r.v. Z .

Prof. Rizzi Laura Tests of Heteroscedasticity


Outline
Introduction
Park test
Goldfeld-Quandt test
White test
Breusch Pagan test

Park test

The Park test is not used frequentely because is not easy to chose the r.v. Z .
When, in cross-section data, the observation units are regions, nations, provinces, etc. the r.v.
(proportionality factor) to be chosen is a size variable which measures indirectly the observational
units dimension.

Example
We use Park test to verify heteroscedasticity in the following data (n = 33):
◮ Y is the number of customers of sampled restaurants;
◮ C is the regressor measuring the number of competitive restaurants;
◮ P is the regressor measuring the resident population;
◮ I is the regressor measuring the average income of resident population.
Estimated equation is:

ŷi = 102, 2 + 9075 Ci + 1288 Ii + 0, 35 Pi


(2053) (0, 54) (0, 073)
−4, 42 2, 37 4, 88

Prof. Rizzi Laura Tests of Heteroscedasticity


Outline
Introduction
Park test
Goldfeld-Quandt test
White test
Breusch Pagan test

Park test

Where R̄ 2 = 0, 58 and F = 15, 75. The estimated auxiliary regression is:

ˆ 2)
ln(e = 21, 05 + 0, 29 ln(Pi )
i
(0, 63)
−0, 46

Given the sample value of the t statistic we accept the null hypothesis (α1 = 0) then accept the
null hypothesis of omoscedasticity.

Prof. Rizzi Laura Tests of Heteroscedasticity


Outline
Introduction
Park test
Goldfeld-Quandt test
White test
Breusch Pagan test

Goldfeld-Quandt test

This test is frequentely used because it is easy to apply when one of the regressors (or another r.v.)
is considered the proportionality factor of heteroscedasticity.
The test has two limits: its difficulty to reject the null hypothesis of omoscedasticity and the fact
that it do not allow to verify other forms of heteroscedasticity.
This test is based on the hypothesis that the error variance is related to a regressor X .
The test procedure is the following:
1 - the observations on Y and X are sorted following the ascending order of the regressor X
which is the proportionality factor;
2 - we divide the sample observations in three subsamples omitting the central one;
3 - we estimate throught OLS the regression models on the first and third subsample (then on
n−c
2 observations each; the number of observations considered has to be sufficiently large);
4 - we calculate the relative RSS, denoted as RSS1 and RSS2 ;
RSS2
5 - we derive the Goldfeld-Quandt test: GQ = R = RSS1 ;
n−c−2k
6 - the test R under the null hypothesis has F distribution with degrees of freedom 2
both for numerator and denominator.

Prof. Rizzi Laura Tests of Heteroscedasticity


Outline
Introduction
Park test
Goldfeld-Quandt test
White test
Breusch Pagan test

Goldfeld-Quandt test

If the sample value of the test F is greater (in a.v.) than the critical value, at the chosen
significance level, we reject the null hypothesis of omoscedasticity.
Idea: if R is large then RSS2 is greater that RSS1 , which means that residuals increase with the
regressor.
The power of this test depends on the number of omitted observations (usually n3 observations
have to be omitted). If we exclude too much observations the RSS2 and RSS1 have too low
degrees of freedom, if we exclude to few observations the test power is low because the comparison
between RSS2 and RSS1 becomes less effective.

Prof. Rizzi Laura Tests of Heteroscedasticity


Outline
Introduction
Park test
Goldfeld-Quandt test
White test
Breusch Pagan test

White test

Sometimes the researcher with to verify if more than one variable is proportionality factor in the
heteroscedasticity process: in these situations it is preferable to consider the Breush Pagan test or
the White test.
The White test has the advantage that it does not assume a specific form of heteroscedasticity.
It is based on a auxiliary regression with suqred residuals as dependent variable and regressors
given by: the regressors of the initial model,, their squares and their cross-products.
The White test procedure is as follows:
1 - we estimate the regression model throught OLS obtaining the OLS residuals, ei . For
instance we estimate: ŷi = b0 + b1 x1i + b2 x2i , then ei = yi − ŷi ;
2 - we estimate an auxiliary regression model with ei2 as dependent variable and initial
regressors, their squares and cross-products as covariates. For instance, we estimate:
ei2 = α0 + α1 x1i + α2 x2i + α3 x1i2 + α4 x2i2 + α5 x1i x2i .
3 - we verify the significance of the auxiliary regression throught the test nR 2 , which, under
the null hypothesis (omoscedasticity) has χ2 (q), where the degrees of freedom q are equal
to the number of regressors in the auxiliary model. In the example q = 5.
4 - if the sample value of the χ2 (q) is greater than the critical one we reject the null
hypothesis of omoscedasticity.

Prof. Rizzi Laura Tests of Heteroscedasticity


Outline
Introduction
Park test
Goldfeld-Quandt test
White test
Breusch Pagan test

White test

This test may have some problems when the number of regressors in the initial model is high.
In these situation the cross-products of regressors may be omitted in the auxiliary model.
When in the initial model there are dummy variables their squares are not included in the auxiliary
regression to avoid multicollinearity problems.

Prof. Rizzi Laura Tests of Heteroscedasticity


Outline
Introduction
Park test
Goldfeld-Quandt test
White test
Breusch Pagan test

Breusch Pagan test

Given the model:


T~
yt = ~
xt β + ut

xtT = [1x2t x3t · · · xkt ].


With t = 1, 2, . . . , n and ~
We assume that heteroscedasticity takes the form:

E (ut ) = 0 for all t


σt2 = E (ut2 ) = h(~ztT α
~)

Where ~ztT = [1z2t z3t · · · zpt ] and α


~ = [α1 α2 · · · αp ] is a vector of unknown coefficients and h(·) is
some not specified function that must take only positive values.
The null hypothesis (omoscedasticity) is then:

H0 : α2 = α3 = · · · = αp = 0

Under the null we have σt2 = h(α1 ) (constant).


The restricted model under the null is estimated throught OLS, assuming distubances normally
distributed.

Prof. Rizzi Laura Tests of Heteroscedasticity


Outline
Introduction
Park test
Goldfeld-Quandt test
White test
Breusch Pagan test

Breusch Pagan test

The test procedure is the following:


xtT ~
◮ estimate the ooriginal model equation by OLS and obtain the OLS residuals, ei = yt − ~ b,
P 2
and the estimated variance of disturbances, σ̃ 2 = et /n;
et2
◮ regress the variable on ~zt by OLS and compute the ESS of this regression;
σ̃ 2
a
◮ under the null hypothesis, H0 we have that: 1 ESS ∼
2 χ2 (p − 1); omoscedasticity is rejected
if 12 ESS exceeds the relative critical value on the χ2 distribution;
◮ a simpler procedure requires the regression of et2 on ~zt ; then the nR 2 of this regression is
asymptotically distributed as a χ2 (p − 1) under the null.

This test needs the knowledge of the regressors ~z but not the knowledge of the functional form
h(·). Sometimes the regressors in ~z may be some regressors included in the original model, in such
case this test becomes an ad hoc version on the White test.

Prof. Rizzi Laura Tests of Heteroscedasticity

You might also like