CHAPTER Assumption Failures

HETEROSCEDASTICITY

Amare Mitiku(PhD), Dep't of Economics,

1
Wollo University
 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 of such violations on estimators?

 How do we detect their presence?

 What are the remedial measures?

Amare Mitiku(PhD), Dep't of Economics,
2
Wollo University
 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.

Amare Mitiku(PhD), Dep't of Economics,

3
Wollo University
 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(ui )   u2 (a constant) but

var(ui )   ui2 (a value that varies)

Amare Mitiku(PhD), Dep't of Economics,

4
Wollo University
Causes for the occurrence of 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.
 Heterosced… 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.
Amare Mitiku(PhD), Dep't of Economics,
5
Wollo University
Consequences of Heteroscedasticity 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 ̂ shall be overestimated.

 As such the t-value associated with it will be underestimated which might lead to the conclusion
that ̂ is statistically insignificant (which in fact may not be true).

 Moreover, our inference and prediction about Dep't

Amare Mitiku(PhD), the population
of Economics, coefficients would be incorrect.
6
Wollo University
 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.
 In Figs. b to e, however, exhibit definite patterns.
Amare Mitiku(PhD), Dep't of Economics,
7
Wollo University
Amare Mitiku(PhD), Dep't of Economics,
8
Wollo University
 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

Amare Mitiku(PhD), Dep't of Economics,

9
Wollo University
a. White test

Amare Mitiku(PhD), Dep't of Economics,

10
Wollo University
b. Breusch-Pagan test

Amare Mitiku(PhD), Dep't of Economics,

11
Wollo University
c. Goldfield-Quandt test

Amare Mitiku(PhD), Dep't of Economics,

12
Wollo University
Amare Mitiku(PhD), Dep't of Economics,
13
Wollo University
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

Amare Mitiku(PhD), Dep't of Economics,

14
Wollo University
Amare Mitiku(PhD), Dep't of Economics,
15
Wollo University
 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.

Amare Mitiku(PhD), Dep't of Economics,

16
Wollo University
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.

Amare Mitiku(PhD), Dep't of Economics,

17
Wollo University
 Ex: Income and Consumption
• We have observations on income and consumption for
30 households.

• We want to test for heteroskedasticity in the following

equation:

Consi   0  1 Inci   i

Amare Mitiku(PhD), Dep't of Economics,

18
Wollo University
 Cont…

• The Goldfeld-Quandt test:

– Rank the observations based on income

– Perform regressions on the first 13 observations
and the last 13 observations.

Amare Mitiku(PhD), Dep't of Economics,

19
Wollo University
 Cont…
• Regression based on first 13 observations
– Cons = 3.4094 + 0.6968Inc
se= (8.7049) (0.0744)
– R2=0.8887 RSS1=377.17 df=11
• Regression based on last 13 observations
– Cons = -28.0272 + 0.7941Inc
se= (30.6421) (0.1319)
– R2=0.7681 RSS2=1536.8 df=11

Amare Mitiku(PhD), Dep't of Economics,

20
Wollo University
 Ex: income and consumption
• Goldfeld-Quandt test statistic:
SSR2 / df 1536.8 / 11
–    4.07
SSR1 / df 377.17 / 11

– F
11; 0.05
11  2.82 ; F 11; 0.01
11  4.46

– Homoskedasticity rejected at 5%-level but not

at 1%-level.

Amare Mitiku(PhD), Dep't of Economics,

21
Wollo University
 Ex: Trade Taxes
• Suppose we have estimated the log-linear model
of the share of trade taxes (import and export)
in total government revenues (y) on a constant,
the share of imports and exports in GNP (x2)
and GNP/capita (x3) for 41 countries.

• How do we perform White’s heteroskedasticity

test?

Amare Mitiku(PhD), Dep't of Economics,

22
Wollo University
 Cont…
• White’s heteroskedasticity test:
– Regress the squared residuals on a constant, x2,
x3, x22, x32 and x2x3
– R2 is 0.1148
– Test statistic: nR2 = 4.7068
– The 5 percent critical chi-square value for 5 df is
11.0705, the 10 percent critical value is 9.2363,
and the 25 percent critical value is 6.62568
– There is no heteroskedasticity.
Amare Mitiku(PhD), Dep't of Economics,
23
Wollo University
MULTICOLLINEARITY

Amare Mitiku(PhD), Dep't of Economics,

24
Wollo University
 The nature of Multicollinearity (M)
 One of the assumptions of the CLRM is that there is no exact
linear relationship exists between any of the explanatory
variables.
 When this assumption is violated, we speak of perfect MC.
Reasons for Multicollinearity
1. The data collection method employed
2. Constraint over the model or in the population being sampled.
3. Over determined model:
 This happens when the model has more explanatory variables
than the number of observations. This could happen in medical
research where there may be a small number of patients about
whom information is collected on a large number of variables.
Amare Mitiku(PhD), Dep't of Economics,
25
Wollo University
 Consequences of Multicollinearity
 If M is perfect, the regression coefficients are indeterminate and
their standard errors are infinite.
 If M is less than perfect (i.e near or high M),
 the regression coefficients are determinate.
 OLS coefficient estimates are still unbiased.
 OLS coefficient estimates will have large variances(or the
variances will be inflated).
 The regression model may do well, that is, R2 may be quite
high.
 Because of the large variance of the estimators, which means
large standard errors, the confidence interval tend to be much
wider, leading to the acceptance of “null hypothesis”.
Amare Mitiku(PhD), Dep't of Economics,
26
Wollo University
 Because of large standard error of the estimators, the
computed t-ratio will be very small leading one or more of the
coefficients tend to be statistically insignificant when tested
individually.
 Although the t-ratio of one or more of the coefficients is very
small (which makes the coefficients statistically insignificant
individually), R2 can be very high.
 The OLS estimators and their standard errors can be
sensitive to small change in the data.

Amare Mitiku(PhD), Dep't of Economics,

27
Wollo University
 Detection of Multicollinearity
 Multicollinearity almost always exists in most applications.
 So the question is not whether it is present or not; it is a
question of degree!
 Also MC is not a statistical problem; it is a data
(sample) problem.
 Therefore, we do not “test for MC’’; but measure its
degree in any particular sample:
1. using variance inflation factor ( for continuous IVs)

Amare Mitiku(PhD), Dep't of Economics,

28
Wollo University
Amare Mitiku(PhD), Dep't of Economics,
29
Wollo University
2. High R2 but few significant t- ratio
3. High pair wise correlation among regressors
4. Test using Eigen values and condition index

Amare Mitiku(PhD), Dep't of Economics,

30
Wollo University
 Remedial measures
 The following corrective procedures have been suggested if
the problem of MC is found to be serious:
1. Increase the size of the sample
2. Introduce additional equation in the model
3. Dropping Variables
4. Transforming variables

Amare Mitiku(PhD), Dep't of Economics,

31
Wollo University
AUTOCORRELATION

Amare Mitiku(PhD), Dep't of Economics,

32
Wollo University
 The nature of Autocorrelation
 In our discussion of simple and multiple regression models,
one of the assumptions of the classicalist is that

 which implies that successive values of the disturbance

term U are independent.
 This means that when observations are made over time, the
effect of disturbance occurring at one period does not carry
over into another period.
 Hence, Autocorrelation is a special case of correlation which
refers to the relationship between successive values of the same
variable.

Amare Mitiku(PhD), Dep't of Economics,

33
Wollo University
Graphical representation of Autocorrelation

Amare Mitiku(PhD), Dep't of Economics,

34
Wollo University
Reasons for Autocorrelation
 There are several reasons why serial or autocorrelation arises.
Some of these are:
a. Cyclical fluctuations
 Time series such as GNP, price index, production, employment
and unemployment exhibit business cycle.
b. Specification bias
This arises because of the following:
i. Exclusion of variables from the regression model
ii. Incorrect functional form of the model
iii. Neglecting lagged terms from the regression model

Amare Mitiku(PhD), Dep't of Economics,

35
Wollo University
Effect of Autocorrelation on OLS Estimators
1. OLS estimates are unbiased
2. The variance of OLS estimate is inefficient
 The variance of the estimates will be biased downwards (i.e.
underestimated) when u’s are auto correlated.
3. Wrong Testing Procedure
 If var(ˆ ) is underestimated, SE ( ˆ ) is also underestimated, this makes t-ratio large.

 This large t-ratio may make ̂ statistically significant while it may not.

4. Wrong testing procedure will make wrong prediction and

inference about the characteristics of the population.

Amare Mitiku(PhD), Dep't of Economics,

36
Wollo University
Detection (Testing) of Autocorrelation
1. Graphic method/Informal Method

Amare Mitiku(PhD), Dep't of Economics,

37
Wollo University
2. Formal testing method
a. Durbin-Watson (DW) test
The DW test statistic is computed as:

d  2(1  ˆ )

Thus we obtain two important conclusions

i. Values of d lies between 0 and 4

ii. If there is no autocorrelation 

ˆ
Amare Mitiku(PhD),
 0, then d  2
Dep't of Economics,
Wollo University
38
 Whenever, therefore, the calculated value of d turns out to be
sufficiently close to 2, we accept null hypothesis(no serial
Correlation), and if it is close to zero or four, we reject the
null hypothesis that there is no autocorrelation.
 However, because the exact value of d is never known,
there exist ranges of values with in which we can either
accept or reject null hypothesis.
 We have dL -lower bound and du - upper bound of the initial
values of d to accept or reject the null hypothesis.

Amare Mitiku(PhD), Dep't of Economics,

39
Wollo University
 Now follow the decision rules given below.

1. If d is less that d L or greater than (4  d L ) we reject the null hypothesis of no autocorrelation in

favor of the alternative which implies existence of autocorrelation.

2. If, d lies between dU and (4  dU ) , accept the null hypothesis of no autocorrelation

3. If however the value of d lies between d L and dU or between (4  dU ) and (4  d L ) , the D.W test
is inconclusive.

Amare Mitiku(PhD), Dep't of Economics,

40
Wollo University
Amare Mitiku(PhD), Dep't of Economics,
41
Wollo University
b. Breusch-Godfrey (BG) Test

Amare Mitiku(PhD), Dep't of Economics, 42

Wollo University
Amare Mitiku(PhD), Dep't of Economics,
43
Wollo University
Remedial Measures for the problems of Autocorrelation
1. If it is pure autocorrelation, one can use appropriate
transformation of the original model so that in the
transformed model we do not have the problem of (pure)
autocorrelation.
 As in the case of heteroscedasticity, we have to use some type
of generalized least-square (GLS) method.
2. In some situations, we can continue to use the OLS method.
3. In large samples, we can use the Newey–West method to
obtain standard errors of OLS estimators that are corrected for
autocorrelation.
 This method is actually an extension of White’s
heteroscedasticity-consistent standard errors method.
Amare Mitiku(PhD), Dep't of Economics,
44
Wollo University
THE END

2O13E.C.
Amare Mitiku(PhD), Dep't of Economics,
45
Wollo University