Download as pdf or txt
Download as pdf or txt
You are on page 1of 19

1

Prof. Monica Roman


1. What is Autocorrelation?
2. SOURCES OF AUTOCORRELATION
3. Consequences
4. Detecting Autocorrelation
5. Correcting Autocorrelation
6. Example
7. References

Prof. Monica Roman 2


 Autocorrelation occurs when assumption IV of the classical LRM
breaks down, meaning that the error term observations in a
regression are correlated

 This phenomenon is common in time series data and causes OLS


estimates to lose some of their nice properties

 In regressions with time series data, the errors associated with


observations typically carry over into future periods

Prof. Monica Roman 3


 Regression model
 Y=X
 Covariance matrix

cov( 1 , 1 ) cov( 1 , 2 )  cov( 1 , n )


cov( 2 , 1 ) cov( 2 , 2 )  cov( 2 , n)

  
cov( n , 1 ) cov( n , 2 )  cov( n , n )

Prof. Monica Roman 4


 The errors are autocorrelated
i j and cov( i, j) 0.

k order correlation coefficient:

cov( i , i k )
k k 1, n - 1
var( i ) var( i k )

Prof. Monica Roman 5


 The typical form of autocorrelation is first-order serial autocorrelation

i i 1 ui
1 1  n 1
 Covariance matrix 1 
2 1 n 2

  
n 1 n 2  1

Prof. Monica Roman 6


 The absence of one or more significant
variables
◦ Example:
yi a bx1i cx 2i i

◦ x3 is omitted the residual variable is explain


through this variable

i x 3i ui

Prof. Monica Roman 7


 Regression model is incorrect specified

 There have been performed inadequate


transformations on data

Prof. Monica Roman 8


 As in the case of heteroskedasticity, autocorrelation leaves the OLS
coefficient estimates unbiased

 However, the OLS estimates do not have the minimum variance (not
efficient estimates)

 Autocorrelation causes OLS to underestimate the standard errors of


the coefficients leading to larger t-statistics and incorrect decisions
in hypothesis testing

Prof. Monica Roman 9


 A standard test of the presence of first-order serial correlation in a
regression model is to examine the residuals from the OLS
estimation

 The Durbin-Watson d statistic uses the regression residuals (e) to


test for first-order serial correlation

n
( ei ei 1 ) 2
i 2
DW n
2
ei
i 1

Prof. Monica Roman 10


 The D-W statistic takes the following “extreme” values

◦ 0 if there is extreme positive serial correlation


◦ 2 if there is no serial correlation
◦ 4 if there is extreme negative serial correlation

 Testing for positive serial correlation, the decision rule is not as


straightforward as in other hypothesis tests because in some cases
the D-W test can be inconclusive

 Critical D-W test statistic values from tables vary with the number of
independent variables and the number of observations

Prof. Monica Roman 11


 To test for positive serial correlation we take the following steps

1. Estimate the model by OLS and obtain the D-W statistic (calculated
by software package)
2. Given the sample size and the number of explanatory variables, find
the upper (dU) and lower (dL) critical values of the d statistic for a
specified level of significance from the table of the D-W statistic
3. Use the following decision rule

◦ 0 < DW < d1 positive serial correlation


◦ d1 DW d2 Inconclusive
◦ d2 < DW < 4-d2 independent
◦ 4-d2 DW 4-d1 Inconclusive
◦ 4-d1< DW <4 positive serial correlation

Prof. Monica Roman 12


Testul Durbin-Watson pentru α= 5 %.
n k=1 k=2 k=3 k=4 k=5
d1 d2 d1 d2 d1 d2 d1 d2 d1 d2
15 1,08 1,36 0,95 1,54 0,82 1,75 0,69 1,97 0,56 2,21
20 1,20 1,41 1,10 1,94 1,00 1,68 0,90 1,83 0,79 1,99
30 1,35 1,49 1,28 1,57 1,21 1,65 1,14 1,74 1,07 1,83
40 1,44 1,54 1,39 1,60 1,34 1,66 1,29 1,72 1,23 1,79
50 1,50 1,59 1,46 1,63 1,42 1,67 1,38 1,72 1,34 1,77
100 1,65 1,69 1,63 1,72 1,61 1,74 1,59 1,76 1,37 1,78

 Note:
◦ d1=dL şi d2=dU

Prof. Monica Roman 13


 1. We estimate the parameters of regression model using
OLS and compute the errors (ei)i=1,n
2. The error term exhibits first-order serial correlation

i i 1 ui n
e i ei 1
i 2
n
ei2 1
i 2

Prof. Monica Roman 14


p
yi 0 j x ji i
j 1
p
yi yi 1 (1 ) ( x ji x ji 1 )
3. 0
j 1
j i i 1

yi* yi yi p
yi* *
1

0 x
j ji i
Having:
x *ji x ji x ji 1
j 1
(1 ) 2
0 0

i N (0, )

4. The parameters of the new model are


estimated

Prof. Monica Roman 15


 The generalized model could be estimated by OLS if we knew the
value of

 One way to estimate the generalized model is the Cochrane-Orcutt


method that eliminates the problem of serial correlation

 The Cochrane-Orcutt method first obtains an estimate of and then


estimates the generalized model by using OLS

 The Cochrane-Orcutt method should, in general, be used, instead of


OLS when autocorrelation is present in a regression model

Prof. Monica Roman 16


Xi yi ei ei-ei-1 (ei-ei-1)2 e 2i

1 10 2 - - 4

2 15 -0,5 -2,5 6,25 0,25

3 20 -3 -2,5 6,25 9

4 30 -0,5 2,5 6,25 0,25

5 40 2 2,5 6,25 4

15 115 - - 25 17,5

Prof. Monica Roman 17


( e i ei 1 ) 2
i 25
DWcalc 1,429
ei2 19,5
i

dL=0,610; dU=1,400.

(du=1,400)<(dcalc=1,429)<(4-1,4),

The errors are not correlated

Prof. Monica Roman 18


 Andrei, T., Bourbonnais, R.- Econometrie, Ed.
Economica, Bucuresti, 2008- capitolul 7, pag.
221-239

 Voineagu, V. si colectiv- Teorie si practica


econometrica, Ed. Meteor Press, 2007, cap.
6.2 pag. 282-294

Prof. Monica Roman 19

You might also like