September 2013
Autocorrelation
September 2013
1/9
Introduction
When dealing with time series data, observations are often correlated across time. Violates the CLR assumption of uncorrelated errors. In this section we look at
1 2 3 4
Autocorrelation
September 2013
2/9
Nature of autocorrelation
Observations in dierent time periods are correlated, i.e. E (ui uj ) 6= 0 for i 6= j . "Autocorrelation" and "serial correlation" Reasons for autocorrelation:
1 2 3 4 5 6 7
Most economic series exhibit business cycles. Omitted variables. Incorrect functional form. Cobweb phenomenon. Data manipulation e.g. smoothing, interpolation, extrapolation Transformation of model e.g. rst dierencing. Non-stationarity in dependent and independent variables.
Autocorrelation September 2013 3/9
What implication is autocorrelation for OLS estimation? Consider the two-variable model:
[6.1] [6.2]
Yt = 1 + 2 Xt + u t ut = ut
1
+ t ,
1<<1
E ( t ) = 0 var (t ) = 2 cov (t , t +s ) = 0, s 6= 0
An error term with the preceding properties is often referred to as a white noise. ut is said to follow a rst-order autoregressive process, or AR(1) process. For 1 < < 1, the error ut is stationary, i.e. constant mean, variance and covariances depend on the time interval between the errors.
HE2004/HE204A Introductory Econometrics () Autocorrelation September 2013 4/9
E ( ut ) = 0
2) = var (ut ) = E (ut 2 1 2
2
s = 0, 1, 2, ....
[6.10]
where xi = Xi
b 2 =
x i y i xi2
X and yi = Yi
2 xi2
[6.11]
var (b 2 ) =
Autocorrelation
September 2013
4/9
[6.12]
2 xt2
var (b 2 )AR 1 =
t xt 1 + 2 x x2 t
t xt 2 + 22 x + x2 t
+ 2n
1 22 x 1 x n xt2
Note the [6.12] is not the same as [6.11], unless = 0. So, usual OLS variance is a biased estimator of the true variance. The true variance [6.12] depends on the sample correlation between the X values, so it is sample dependent. In general, it is not possible to tell if var (b 2 ) > var (b 2 )AR 1 or var (b 2 ) < var (b 2 )AR 1 . Hence, inference based on [6.11] would be incorrect. What about the linear unbiased property of OLS? Under autocorrelation OLS estimator continues to be linear and unbiased. But it is not "best", ie. it is not e cient. Within the class of linear unbiased estimators, there is a more e cient estimator.
HE2004/HE204A Introductory Econometrics () Autocorrelation September 2013 5/9
[6.13]
Yt
= 1 + 2 Xt
+ ut
[6.14]
Yt
Yt ut
1 1
= (1
) 1 + 2 ( Xt
Xt
1 ) + vt
[6.16]
E ( v t ) = E ( ut
ut
1)
=0
[6.17] [6.18]
Equation [6.17] is not feasible since usually we do not know . A feasible version of it involves estimating which we will discuss later.
GLS b 2 =
1)
Autocorrelation
September 2013
7/9
2 / (n b 2 = u The residual variance bt 2) is likely to underestimate the 2 true variance . When there is AR(1) in the error term
[6.19]
1 n 2 where r = n t =1 xt xt 1 / t =1 xt , which can be interpreted as the sample correlation coe cient between successive values of the X 0 s .
HE2004/HE204A Introductory Econometrics () Autocorrelation September 2013 8/9
b2 ) = E (
2 fn [2 /(1 )] 2 r g n 2
Another source of downward biasedness of the estimate of var (b 2 ) can be seen by comparing [6.11] and [6.12]. If > 0 and the X 0 s are positively correlated, then var (b 2 ) < var (b 2 )AR 1 . Hypotheses testing based on var (b 2 )AR 1 , while valid, generally lacks power.
b2 ) < 2 , i.e. the usual residual If both > 0 and r > 0, then E ( variance will underestimate the true variance. This will then lead to downward bias of the estimate of var (b 2 ) since we typically estimate 2 2 b /xt . it by
Autocorrelation
September 2013
9/9