ST 4011 – Econometrics (30L, 2C) Therefore, the variance of the disturbance terms 𝑢𝑖 ; 𝑖 = 1, 2, 3, … , 𝑛 do

Department of Statistics
not take some constant value 𝜎 2 but the values are varied depending on
the observed values of the explanatory variable 𝑋. Figure 01 illustrates a
Heteroscedasticity
heteroscedasticity situation with unequal error variances.
The underlying idea behind the heteroscedasticity is that the disturbance
or the error terms of a regression model to have unequal variances. The
variance of each disturbance term 𝑢𝑖 conditional on the chosen values of
explanatory variable(s) to take some constant value 𝜎 2 is one of the
important assumptions of the classical linear regression models.

Consider the following simple linear regression model

𝑌 = 𝛽0 + 𝛽1 𝑋 + 𝑢
that can be written for the 𝑖 𝑡ℎ observation of a random sample of size 𝑛 as
𝑦𝑖 = 𝛽0 + 𝛽1 𝑥𝑖 + 𝑢𝑖 ; for 𝑖 = 1, 2, 3, … , 𝑛.
Here, 𝑦𝑖 and 𝑢𝑖 are random variables and 𝑥𝑖 is assumed fixed within the
Figure 01 – Heteroscedasticity situation with unequal error variances
observed range of values of the 𝑋 variable in the sample.

Under constant variance or homoscedasticity condition,

𝑉(𝑦𝑖 |𝑋 = 𝑥𝑖 ) = 𝑉(𝑢𝑖 |𝑋 = 𝑥𝑖 ) = 𝐸(𝑢𝑖2 |𝑋 = 𝑥𝑖 ) = 𝜎 2 .
Note that the conditional variance of 𝑦𝑖 remains same regardless of the
values taken by 𝑥𝑖 . However, in the case of heteroscedasticity,
𝑉(𝑦𝑖 |𝑋 = 𝑥𝑖 ) = 𝑉(𝑢𝑖 |𝑋 = 𝑥𝑖 ) = 𝐸(𝑢𝑖2 |𝑋 = 𝑥𝑖 ) = 𝜎𝑖2 .
Department of Statistics ST 4011 - Econometrics

There are several reasons for the heteroscedasticity problem of the error 3. Improvements in data collection techniques can cause to reduce the
terms in practice. Some of the reasons are as follows: error variance of outcomes.
1. Error-learning models Example: Banks that have sophisticated data processing equipments
As people learn, their errors of behavior become smaller over time. are likely to commit fewer errors in the monthly or quarterly
Therefore, the variance of the outcome of the learning process can statements of their customers.
expect to reduce over time. 4. Heteroscedasticity can also arise due to the effects of outlier
Example: For a typist, as the number of hours of typing practice observations.
increases, the average number of typing errors and the variance of the An outlier is an observation that is much different to the remaining
number of errors can expect to be decreased over time. observations in a dataset (either considerably large or small compared
2. As incomes grow, people have more discretionary income and to the others). Therefore, the inclusion of such outlier observations,
therefore more options will be available to spend their income. Hence, especially if the sample size is small, can cause to violate the
the variance of savings is likely to be increased with income. homoscedasticity requirement in linear regression models.
Note: The discretionary income refers to the income remaining after 5. Heteroscedasticity may appear due to the omission of important
the deduction of taxes, social security charges, and other basic living variables from a model.
costs. Consider the following model
Similarly, the companies with larger profits tend to show greater 𝑌 = 𝛽0 + 𝛽1 𝑋1 + 𝛽2 𝑋2 + 𝑢
variability in their dividend (payment) policies than the companies with Suppose that the following model has been fitted by ignoring 𝑋2
lower profits. When a company makes a profit, it is required to take variable.
decision whether to retain the profit in the company or distribute the 𝑌 = 𝛽0 + 𝛽1 𝑋1 + 𝑣
money to shareholders in the form of dividends. Dividend policy
indicates the amount of dividends paid out by the company to its
shareholders and the frequency with which the dividends are paid out.

2
Department of Statistics ST 4011 - Econometrics

Then, 𝑣 is a function of the ignored 𝑋2 . Hence, 𝑣 changes as the value 8. Incorrect data transformations
of 𝑋2 changes. Therefore, cannot expect to hold the constant variance Example: Ratio transformation
assumption of the error terms 𝑣 in the reduced model. Please refer to the following linear regression model that was
6. Heteroscedasticity can also appear due to the skewness of the discussed in the previous lesson
distributions of one or more explanatory variables in linear regression 𝑌𝑡 1 𝑋1𝑡 𝑢𝑡
= 𝛽0 ( ) + 𝛽1 ( ) + 𝛽2 +
𝑋2𝑡 𝑋2𝑡 𝑋2𝑡 𝑋2𝑡
models.
as a remedy for the multicollinearity problem in the original model
Example: It is a well-known fact that the economic variables such as
𝑌𝑡 = 𝛽0 + 𝛽1 𝑋1𝑡 + 𝛽2 𝑋2𝑡 + 𝑢𝑡 ;
income and wealth in most societies are uneven with the bulk of
income and wealth being owned by few at the top. Therefore, if we where
use these types of variables with skewed distributions as the
𝑌𝑡 − Consumption expenditure at time 𝑡
explanatory variables in linear regression models, then the constant
variance assumption of errors can generally be violated. 𝑋1𝑡 − Gross Domestic Product (GDP) at time 𝑡
7. Incorrect functional forms
𝑋2𝑡 − Total population at time 𝑡.
Example: Linear vs log-linear models
The log-linear transformation can be applied to reduce the spread of Note: The heteroscedasticity problem is more common in cross-sectional
data and therefore to satisfy the constant variance assumption of than time series data. In cross-sectional data, studies are conducted on
errors in linear regression models. Therefore, if we ignore the population units at a given point in time. For example, population units can
potential log-linear model in a suitable context and apply the linear be individual consumers or their families, firms, industries, geographical
model instead, then the heteroscedasticity problem can appear. subdivisions such as countries, states, and cities and so on. These units may
be nonhomogeneous due to various reasons such as different in sizes
leading to heteroscedasticity problem in regression analyses.

3
Department of Statistics ST 4011 - Econometrics

In time series data, on the other hand, aggregated data are obtained from and
the same entity over a period of time such as gross national product (GNP), ∑𝑛 2 2
𝑖=1(𝑥𝑖 −𝑥̅ ) 𝜎𝑖
𝑉 (𝛽̂1 ) = 2 (under heteroscedasticity).
[∑𝑛 2
𝑖=1(𝑥𝑖 −𝑥̅ ) ]
total consumption expenditure, total savings, and total employment and
so on. Therefore, the variability of observations may not be considerably
Generalized Least Square (GLS) Method
varied over time.
In this approach, observations that are coming from populations with
greater variability are given less weights than those are coming from
OLS Estimation in the presence of heteroscedasticity
populations with smaller variability. The OLS method does not consider the
Consider the following simple linear regression model is written for the 𝑖 𝑡ℎ
information that the dependent variable 𝑌 has non-constant variability. It
observation in a sample of size 𝑛 as discussed above.
assigns equal weights or importance to each observation. But the GLS
𝑦𝑖 = 𝛽0 + 𝛽1 𝑥𝑖 + 𝑢𝑖 ; 𝑖 = 1, 2, 3, … , 𝑛.
method takes that information into account and assigns different weights
Suppose we apply the OLS method in the presence of heteroscedasticity.
for observations depending on their variability leading to obtain best linear
Under heteroscedasticity, 𝑉(𝑢𝑖 ) = 𝜎𝑖2 . Therefore, 𝐸(𝑢𝑖 2 ) = 𝜎𝑖2
unbiased estimators (BLUE) in practice.
(∵ 𝐸(𝑢𝑖 ) = 0)
Consider the above used regression model 𝑦𝑖 = 𝛽0 + 𝛽1 𝑥𝑖 + 𝑢𝑖
Let us assume that all the other assumptions of the classical linear model
and write it as
are satisfied.
𝑦𝑖 = 𝛽0 𝑥0𝑖 + 𝛽1 𝑥1𝑖 + 𝑢𝑖
It can be shown that the OLS estimators are still unbiased (as there is no
where 𝑥0𝑖 = 1 for each 𝑖 for the ease of algebraic manipulation.
impact of the heteroscedastic condition on the unbiasedness property).
Suppose that the heteroscedastic variances 𝜎𝑖2 𝑠 are known and obtain the
However, the OLS estimators do not have minimum variance. Therefore,
following equation by dividing the above equation by 𝜎𝑖
we cannot say that they are efficient. Some adjustment should be made to
𝑦𝑖 𝑥0𝑖 𝑥1𝑖 𝑢𝑖
the model to produce estimators that are BLUE. = 𝛽0 + 𝛽1 + .
𝜎𝑖 𝜎𝑖 𝜎𝑖 𝜎𝑖

For example, it can be shown that For the ease of explanation, it can be written as
𝜎2 𝑦𝑖 ∗ = 𝛽0 ∗ 𝑥0𝑖 ∗ + 𝛽1 ∗ 𝑥1𝑖 ∗ + 𝑢𝑖 ∗ .
𝑉 (𝛽̂1 ) = ∑𝑛 2
(under homoscedasticity)
𝑖=1(𝑥𝑖 −𝑥̅ )

4
Department of Statistics ST 4011 - Econometrics

Now consider the variance of 𝑢𝑖 ∗ . ∑𝑛

𝑖=1 𝑤𝑖
and 𝑉(𝛽̂1∗ ) = 2; where 𝑤𝑖 = 1⁄ ,
2 [(∑𝑛 𝑛 2 𝑛
𝑖=1 𝑤𝑖 )(∑𝑖=1 𝑤𝑖 𝑥1𝑖 )]−[∑𝑖=1 𝑤𝑖 𝑥1𝑖 ]
𝜎𝑖2
∗2 𝑢
V (𝑢𝑖 ∗ ) = 𝐸[𝑢𝑖 ] = 𝐸 [(𝜎𝑖) ]
𝑖 respectively.
1
= 𝐸(𝑢𝑖2 )
𝜎𝑖2
1
= 𝜎2 𝜎𝑖2 Since E (𝑢𝑖2 ) = 𝜎𝑖2 as E (𝑢𝑖 ) = 0
𝑖

=1 (Constant and hence, homoscedasticity) OLS – Same weights are given

Now the variance of the transformed disturbance term 𝑢𝑖 ∗ is GLS – Lesser weight is given to
observation C with larger variance
homoscedastic. Hence, if we apply the OLS method on the transformed
model, the resulting estimators are BLUE.

𝑦𝑖 ∗ = 𝛽0 ∗ 𝑥0𝑖 ∗ + 𝛽1 ∗ 𝑥1𝑖 ∗ + 𝑢𝑖 ∗

Here, GLS estimators can be derived by minimizing ∑𝑛𝑖=1 𝑢𝑖 ∗ 2 . Note that

Figure 02: OLS and GLS weights
∑𝑛𝑖=1 𝑢𝑖 ∗ 2 = ∑𝑛𝑖=1(𝑦𝑖∗ − 𝛽0 ∗ 𝑥0𝑖 ∗ − 𝛽1 ∗ 𝑥1𝑖 ∗ )2
𝑛
𝑦𝑖 𝑥0𝑖 𝑥1𝑖 2
= ∑( − 𝛽0 ∗ − 𝛽1 ∗ ) Note: In GLS the weight assigned to each observation is inversely
𝜎𝑖 𝜎𝑖 𝜎𝑖
𝑖=1
proportional to 𝜎𝑖2 . Hence, observations that are coming from populations
with larger variances will get smaller weights. Under GLS, observation C will
Exercise: Show that the corresponding estimators of 𝛽̂1∗ and 𝑉(𝛽̂1∗ ) are
get relatively lower weight than the other two observations A and B in the
[(∑𝑛𝑖=1 𝑤𝑖 )(∑𝑛𝑖=1 𝑤𝑖 𝑥1𝑖 𝑦𝑖 )] − [(∑𝑛𝑖=1 𝑤𝑖 𝑥1𝑖 )(∑𝑛𝑖=1 𝑤𝑖 𝑦𝑖 )]
𝛽̂1∗ = above figure 02. This approach allocates higher weights to observations
[(∑𝑛𝑖=1 𝑤𝑖 )(∑𝑛𝑖=1 𝑤𝑖 𝑥1𝑖2 )]
− [∑𝑛𝑖=1 𝑤𝑖 𝑥1𝑖 ]2
that are closely clustered around their population means.