if λt is random, and {μi } to be spatially correlated if μi is random. We also allow a certain
degree of heterogeneity in panel data—{Yn , X1ii}' need not be stationary for each i, and the
errors {Vii} may have different variances across i. In particular, we allow some nonstationary
processes. One example of nonstationary panel time series is the deterministic trend process
(e.g., Kao and Emerson 1999)
Yt = ɑ + Ylw + 72w2 + ’ ’ ’ + Ypwp + μi + λ + vit =
This is covered by (2.1) with Xit ≡ [t/т, (t∕T)2, ...> (t∕T)p]z and θ ≡ (Tγ1, ...,Tpγp)z. Note
that Xn and β0 depend on T. Another example is the panel cointegration process (e.g.,
Phillips and Moon 1999, Kao and Chiang 2000):
Yit = ɑ + γZit + μi + λt + Va,
where Zn = Zn-ɪ + %n, {ε* j is I(0) for each i, and {%i j may or may not be correlated with
{Vi j. This process is also covered by (2.1) with Xit ≡ T~1Zn and β0 ≡ Tγ. We will provide
regularity conditions on transformed variables {Xi j and transformed parameters β.
The parameter vector β0 in (2.1) can be estimated by the popular within estimator
∑ ∑ (X« - v - v' X) (Xft - X - Xt + X)z
i=l t=l
■ n Ti
∑ ∑ ( Xn - Xi - Xt + X ) ( Yt - Y - Y + γ )
i=l t=l
(2.2)
where Xt ≡ Ti 1 S/ɪ Xtt, Xt ≡ n 1 S"= ɪ Xit and X ≡ (nTi) 1 S"=1 S’ɪ Xit. The variables
Yi,Yt and Y are defined in the same ways. Its asymptotic covariance matrix is
-ι / ʌ ∖-ɪ
p Iim lnT Ip Iim MnT\ , (2.3)
. . i-------- ,'^T _ , / . ʌ ∖
AVAR TTY (β - βfi ) = (p Iim MnT]
where
1 n τi
Mτ≡ Tf ∑∑ Xv-
n
i=l t=l
• n Ti Ti
lnT ≡ ~TTt ΣΣΣXitVit(visXis') ,
П i=l t=l s=l
Xn ≡ Xn - X7i - Xt + X7, and Vn ≡ Vn - Vβ - vt + v. To estimate (2.3), pIim Ωnr is more
challenging to estimate.
More generally, for many panel estimators β, we have
(MnτΩnM'nτ)-1 YTt(e - β0) $ N(0, Ir), u ∈ Z+,
(2.4)