Scribd
Upload
261 views

Lecture Notes GLS

This document discusses methods for correcting heteroskedasticity and autocorrelation when estimating regression models. It first describes possible causes and effects of heteroskedasticity and autocorrelation. To correct heteroskedasticity, the document proposes using weighted least squares where observations are weighted by the inverse of the estimated variance. To correct autocorrelation, it proposes using generalized least squares by transforming the model using lagged dependent and independent variables. Both of these methods aim to produce parameter estimates that are best linear unbiased by satisfying the assumptions of the classical linear regression model.

Uploaded by

Kasem Ahmed
Copyright
© © All Rights Reserved
Available Formats
Download as DOCX, PDF, TXT or read online on Scribd
261 views

Lecture Notes GLS

This document discusses methods for correcting heteroskedasticity and autocorrelation when estimating regression models. It first describes possible causes and effects of heteroskedasticity and autocorrelation. To correct heteroskedasticity, the document proposes using weighted least squares where observations are weighted by the inverse of the estimated variance. To correct autocorrelation, it proposes using generalized least squares by transforming the model using lagged dependent and independent variables. Both of these methods aim to produce parameter estimates that are best linear unbiased by satisfying the assumptions of the classical linear regression model.

Uploaded by

Kasem Ahmed
Copyright
© © All Rights Reserved
Available Formats
Download as DOCX, PDF, TXT or read online on Scribd
You are on page 1/ 5

Econometrics Lecture Notes

Estimation Methods in Cases of Heteroskedastic or Autocorrelated


Disturbance Terms

Possible Causes of Heteroskedasticity:


1. The values of the variables in the regression equation vary substantially in different
observations. The variation in the omitted variables and the measurement errors which
are jointly responsible for the disturbance term will be relatively small when y and x are
small and relatively large when they are large (economic variables tending to move in
size together).
2. In time series analysis if y and x are growing over time, then it may well happen that the
variance of the disturbance term is also growing over time.

II Effects of Heteroskedasticity:
1. OLS is no longer efficient, since it treats all observations as being of equal importance,
while there is much more information about the line to be had from some observations
than from others, and when estimating the line more attention should be paid to the
observations having small variances than those with larger variances.
2. The standard errors are biased, and thus hypothesis tests and confidence intervals based
on the t-distribution are likely to be invalid.

III Correcting Heteroskedasticity (Weighted Least Squares)


The way to correct heteroskedasticity is to postulate a model for the disturbances, and on the
basis of this model seek transformations of the data so that with the transformed data OLS is
an efficient procedure.

Let us assume that the variance of (t ) is directly proportional to the variable Z (t ) and the model
for heteroskedasticity is

2 2 Z (t ) ,

(1)

2
where is an unknown parameter and Z (t ) is a variable on which we have observations.

The full heteroskedastic model may now be written as


y (t ) a bx (t ) (t ) ,

(2)

where
E ( (t )) 0,

E ( 2 (t )) 2( t ) 2 Z (t ),
and E ( (t ), (t )) 0 when t .
Now divide equation (2) by

Z (t ) to obtain

y (t )
1
x(t )
a
b
,
Z (t )
Z (t )
Z (t )

(3)

(t )
~ (t )
.
Z (t )
where
If we introduce the notation
y (t )
~
y (t )
, xo (t )
Z (t )

1
x (t )
,~
x (t )
,
Z (t )
Z (t )

then the transformed model is

~y (t ) a ~
xo (t ) b ~
x (t ) ~(t ).
The OLS estimates of the unknown parameters a and b in the transformed model are BLUE,
because the disturbance term of the transformed model satisfies the classical disturbance
assumptions:

(4)

(a)

(t )

E (~ (t )) E

Z
(
t
)

1
E ( (t ))
Z (t )

(because Z(t) is not stochastic) = 0 (because

E ( (t )) 0),

(b)

(c)

2 (t )
1
1

E (~ 2 (t )) E
E ( 2 (t ))
2 Z (t ) 2 ,
Z (t )
Z (t )
Z (t )

(t )
(t )
1
E (~ (t ), ~ (t )) E

Z (t ) Z (t ) E ( (t ), (t )) 0.
Z
(
t
)
Z
(
t

Note: dividing by

Z (t ) actually means weighting the observations according to their relative

importance ( the estimating procedure is termed weighted least squares).

IV Possible Causes of Residual (Serial) Autocorrelation:


Autocorrelation normally occurs in regression analysis using time series data. The disturbance

(t ) may be regarded as summarizing all the influences on Y not accounted for by the
explanatory variables X. As successive values of economic variables are nearly always highly
correlated, it is reasonable to assume that the disturbances are autocorrelated.

V Main Effects of Autocorrelation:


1. A loss of efficiency. If disturbances are correlated, then previous values of the
disturbance have some information to convey about the current disturbance. If this
information is ignored, it is clear that the sample data are not being used with maximum
efficiency.
2. If there is a positive autocorrelation, the standard errors of the coefficients will be biased
downward. In the results from a particular sample, this is likely to show up in confidence
2
intervals that are much narrower than they should be and values of R that give an overly

optimistic impression of the worth of the model for prediction. Parameter estimates
associated with irrelevant variables may be highly significant.

VI Overcoming Autocorrelation (Generalized Least Squares)


Suppose that y (t ) a b x (t ) (t ),

(5)

where

(t ) (t 1) u (t ).

(6)

The value of u in each observation is assumed to be independent of its value in all other
observations. Assume for the moment that is known, although in practice this is never the case.
Our strategy is to transform (5) so that the disturbance for the transformed model has no
autocorrelation. OLS is then used on the new model to produce parameters estimators that are
BLUE.
We lag (5) by one period and multiply by :

y (t 1) a b x (t 1) (t 1), t 2, T .

(7)

Subtracting (7) from (5) produces


y (t ) y (t 1) a (1 ) b ( x (t ) x (t 1)) (t ) (t 1),

(8)

and since from (6) we may write


u (t ) (t ) (t 1),

(9)

equation (8) can be rewritten as


y (t ) y (t 1) a (1 ) b ( x (t ) x (t 1)) u (t ).
(10)
If we introduce the notation

~y (t ) y (t ) y (t 1), x (t ) 1 , ~
x (t ) x(t ) x (t 1),
o
then the transformed model is

~y (t ) a ~
xo (t ) b ~
x (t ) u (t ).

(11)

The disturbance term for this model was originally assumed to exhibit no autocorrelation. But
the number of observations to be used for fitting the line is only T-1 instead of T. This might
cause the loss of efficiency that overweighs the gain due to overcoming autocorrelation.
2
Therefore it may be useful to regain the first observation y(1), x(1) with its residual variance .

The problem is that the latter is different from the residual variance in the remaining
2
observations, u . So we have solved the problem of autocorrelation at the expense of

introducing a special case of heteroskedasticity.


Using (6), we can express

2 in terms of u2 :

2 var( ) var( (t 1) u (t )) 2 var( (t 1)) var( u (t )) 2 cov( (t 1), u (t ))


2 2 u2 ,

(12)

2
since (t 1) and u(t) are independent and since var (t ) var( (t 1)) . Hence

u2
.
1 2

(13)

2
This allows us to eliminate heteroskedasticity by weighting the 1st observation by 1 .

2
The disturbance term in that observation becomes 1 (1). Its variance is

2 (1 2 ) 2 u2 (see (13)).

So after regaining the first observation and eliminating heteroskedasticity we obtain a classical
linear regression model, which means that applying OLS will produce parameters estimates
which are BLUE (Best Linear Unbiased Estimates).

You might also like