Appendix to “Asymptotically unbiased estimation of
autocovariances and autocorrelations for panel data with
incidental trends”
Ryo Okui
∗February 3, 2011
1 Proof of Theorem 1
Let xt= (1, t)′, βi= (ηi, δi)′and ˆβi= (ˆηi, ˆδi)′. Then we can write ˜yit= yit−x′tβˆi = x′t(βi− ˆβi)+wit, and:
ˆ
γk = 1
N (T − k)
N
∑
i=1 T
∑
t=k+1
˜ yity˜i,t−k
= 1
N (T − k)
N
∑
i=1 T
∑
t=k+1
witwi,t−k+ 1 N (T − k)
N
∑
i=1 T
∑
t=k+1
x′t(βi− ˆβi)wi,t−k
+ 1
N (T − k)
N
∑
i=1 T
∑
t=k+1
x′t−k(βi− ˆβi)wit+ 1 N (T − k)
N
∑
i=1 T
∑
t=k+1
(βi− ˆβi)′xtx′t−k(βi− ˆβi).
Lemma 1 in Okui (2010) gives: 1 N (T − k)
N
∑
i=1 T
∑
t=k+1
witwi,t−k→pγk
and
√N T
( 1
N (T − k)
N
∑
i=1 T
∑
t=k+1
witwi,t−k− γk
)
→dN
0,
∞
∑
j=−∞
{γj2+ γk+jγk−j+ cum(0, −k, j, j − k)}
.
Next, we consider the second term in the expansion of ˆγk. We have: 1
N (T − k)
N
∑
i=1 T
∑
t=k+1
x′t(βi− ˆβi)wi,t−k = −
1 N
N
∑
i=1
1 T − k
( T
∑
t=k+1
xtwi,t−k
)′( T
∑
t=1
xtx′t
)−1( T
∑
t=1
xtwi,t
) .
∗Institute of Economic Research, Kyoto University, Yoshida-Hommachi, Sakyo, Kyoto, Kyoto, 606-8501, Japan. Email: okui@kier.kyoto-u.ac.jp
1
Simple matrix algebra shows that: ( T
∑
t=k+1
xtwi,t−k
)′( T
∑
t=1
xtx′t
)−1( T
∑
t=1
xtwi,t
)
= 4 T
( T
∑
t=1
wit
) (T −k
∑
t=1
wit
)
−T (T + 1)6 ( T
∑
t=1
twit
) (T −k
∑
t=1
wit
)
−T (T + 1)6 ( T
∑
t=1
wit
) (T −k
∑
t=1
twit
)
+ 12
T (T + 1)(T − 1) ( T
∑
t=1
twit
) (T −k
∑
t=1
twit
) .
Let ¯wi =∑Tt=1wit/T and ¯zi =∑Tt=1twit/T2. Then, because ¯wi = Op(1/√T ), ¯zi = Op(1/√T ) and wit’s are i.i.d. over individuals and have finite fourth moments, we have that:
1 N
N
∑
i=1
1 T − k
( T
∑
t=k+1
xtwi,t−k
)′( T
∑
t=1
xtx′t
)−1( T
∑
t=1
xtwi,t
)
= 1 N
N
∑
i=1
T T − k(4 ¯w
2
i − 12 ¯wiz¯i+ 12¯zi2) + O ( 1
T3/2 )
.
We now note that:
E( ¯w2i) = γ0+ 1 T
T −1
∑
j=0
2T − j T γj=
VT
T ,
E( ¯wiz¯i) = 1 T3
T
∑
t=1
t
T
∑
j=1
γ|t−j| =1 2
VT
T + o ( 1
T )
,
E(¯zi2) = 1 T4
T
∑
t=1
t
T
∑
j=1
jγ|t−j| =1 3
VT
T + o ( 1
T )
.
Therefore, we have that:
E
1 N
N
∑
i=1
1 T − k
( T
∑
t=k+1
xtwi,t−k
)′( T
∑
t=1
xtx′t
)−1( T
∑
t=1
xtwi,t
)
= 4VT T − 12
1 2
VT
T + 12 1 3
VT
T + o ( 1
T )
= 2VT T + o
( 1 T
) .
Moreover, because wit have finite fourth moments by assumption, we have var( ¯w2i) = O(T12) as well as var( ¯wi¯zi) = O(T12) and var(¯zi2) = O(T12). It therefore follows that:
var
1 N
N
∑
i=1
1 T − k
( T
∑
t=k+1
xtwi,t−k
)′( T
∑
t=1
xtx′t
)−1( T
∑
t=1
xtwi,t
)
= o ( 1
N T )
.
Therefore, we have that: 1
N
N
∑
i=1
1 T − k
( T
∑
t=k+1
xtwi,t−k
)′( T
∑
t=1
xtx′t
)−1( T
∑
t=1
xtwi,t
)
= 2
TVT + op ( 1
√N T )
,
and
1 N (T − k)
N
∑
i=1 T
∑
t=k+1
x′t(βi− ˆβi)wi,t−k= −2
TVT+ op ( 1
√N T )
,
2
when N/T → c. Similarly, we have:
1 N (T − k)
N
∑
i=1 T
∑
t=k+1
x′t−k(βi− ˆβi)wit= −2
TVT + o ( 1
√N T )
,
and
1 N (T − k)
N
∑
i=1 T
∑
t=k+1
(βi− ˆβi)′xtx′t−k(βi− ˆβi) = 2 TVT + o
( 1
√N T )
.
To sum up, we get:
√N T (
ˆ
γk− γk+ 2 TVT
)
→dN
0,
∞
∑
j=−∞
{γ2j+ γk+jγk−j+ cum(0, −k, j, j − k)}
.
References
Okui, R. (2010). Asymptotically unbiased estimation of autocovariances and autocorrelations with long panel data, Econometric Theory 26: 1263–1304.
3