CHAPTER 4
HETEROSCEDASTICITY
Introduction
In both the simple and multiple regression models, we made
assumptions.
Now, we are going to address the following ‘questions:
What if the error variance is not constant over all observations?
What if the different errors are correlated?
What if the explanatory variables are correlated?
What are the consequences such violations on estimators?
How do we detect their presence?
What are the remedial measures?
In general we could encounter any combination of 3
problems:
the coefficient estimates are wrong
the associated standard errors are wrong
the distribution that we assumed for the test statistics will
be inappropriate.
Heteroscedasticity
The nature of Heteroscedasticty
The assumption of homoscedasticity states that the
variation of each ui around its zero mean does not depend on
the value of explanatory variable.
Mathematically, u2 is not a function of X; i.e. 2 f ( X i )
When the variance of the error term is different for
different values of X you have heteroskedasticity.
Thus, we say that U’s are heteroscedastic when:
var( u i ) u2 (a constant) but
var( u i ) ui2 (a value that varies)
Reasons for Hetroscedasticity
Error learning model:
It states that as people learn, their error of behavior
become smaller over time.
In this case is expected to decrease.
As data collection technique improves, is likely to decrease.
H can also arise as a result of the presence of outliers.
An outlier is an observation that is much different
(either very small or very large) in relation to the other
observation in the sample.
Consequences of Hetroscedasticity for the OLS estimates
The OLS estimators will have no bias
i.e., the estimates are unbiased even under the condition of H.
Variance of OLS coefficients will be incorrect
the estimators do not have the smallest variance in the class of
unbiased estimators and, therefore, they are not efficient.
var( ˆ ) under heteroscedasticty will be greater than its variance under homoscedasticity.
As a result the true standard error of ˆ shall be overestimated.
As such the t-value associated with it will be underestimated which might lead to the conclusion
that ˆ is statistically insignificant (which in fact may not be true).
Moreover, our inference and prediction about the population coefficients would be incorrect.
Detecting Heteroscedasticity
two methods of testing or detecting heteroscedasticity.
i. Informal method
ii. Formal method
i. Informal method
informal because it does not undertake the formal testing
procedures such as 2-test, F-test and the like.
whether a given data exhibits H or not, we look on whether
there is a systematic relation between residual squared e2i
and the Predicted value of Y or Xi.
In fig (a), we see there is no systematic pattern between the
two variables, suggesting that no H is present in the data.
Figures b to e, however, exhibit definite patterns.
Visual Inspection/graphical
Do a visual inspection of residuals plotted against fitted values;
or, plot the IV suspected to be correlated with the variance of the
error term.
In Stata, after running a regression, you could use the rvfplot
(residuals versus fitted values) or rvpplot command
(residual versus predictor plot, e.g. plot the residuals versus one
of the X variables included in the equation) or
Statistics ► linear models and related ► regression
diagnostics ► residual vs fitted or residual vs predictor plot.
In SPSS, plots could be specified as part of the Regression
command:
Analyze ► regression ► linear ► plots ► ZRESID(y-axis) vs
ZPRED(x-axis) ► ok
ii. Formal methods
several formal methods of testing heteroscedasticty which are
based on the formal testing procedures.
some of the major ways of detecting heterosedasticity.
a. White test
b. Breusch – Pagan Test
c. Goldfield-Quandt test
a. Breusch-Pagan test
Breusch-Pagan / Cook-Weisberg Test for Heteroskedasticity
The test is designed to detect any linear form of H.
You run a regression, and then give the estat hettest command (or,
hettest alone will work) or
Statistics ► linear models and related ► regression
diagnostics ► specification tests, etc ► hettest
A large chi-square would indicate that H was present.
i.e. if there is H, then the test statistic will be significant.
b. White test
White’s General Test for H
The default Breusch-Pagan test specified by hettest is a test for
linear forms of H.
In this default form, the test does not work well for non-linear
forms of H.
The default test also has problems when the errors are not
normally distributed.
This can be estimated via the command estat imtest, white or
just imtest, white or imtest only or
After OLS regression, go to Statistics ► linear models and related
► regression diagnostics ► specification tests, etc ► Information
matrix test (imtest)
is a special case of the Breusch-Pagan test, where the
assumption of normally distributed errors has been also
relaxed (Skewness).
c. Goldfeld-Quandt test
Remedial measures for the problems of heteroscedasticity
H does not destroy the unbiasedness of the OLS estimators,
but they are no longer efficient.
remedial measures concentrate on the variance of the error
term.
1. Transforming the model
In short GLS is OLS on the transformed variables that satisfy the standard least squares assumptions.
The estimators thus obtained are known as GLS estimators, and it is these estimators that are
BLUE.
2. Taking a robust standard error during regression
how to adjust standard errors, t and F statistics so that they
are valid in the presence of H of unknown form.
it means we can report new statistics that work, regardless
of the kind of H present in the population.
The method is known as heteroskedasticity-robust
procedures because they are valid—at least in large
samples—whether or not the errors have constant
variance.
Once H-robust standard errors are obtained, it is simple to
construct a H-robust t statistic.