where W(s) is a d-dimensional standard Brownian motion with d = K - r. Note that (2) is
the limiting null distribution of the trace statistic (LR statistic) for a given individual i, i.e.
of -T jK=r+1 ln(1 - bλij); see, e.g., Johansen (1995).
Under the null hypothesis and assuming suitable conditions, Larsson et al. (2001) applied
a central limit theorem to establish the asymptotic normality of their standardized LR-bar
statistic, so that standard normal quantiles may serve as critical values for the test. Moreover,
they approximated the first two moments of the asymptotic trace statistic Zd for different
values d by simulation as sample moments of
ZT,d = tr
1 T 1 T -1
T ∑εtX0-1 T2 Σx<-1 Xt-1
t=1 t=1
1T
T X Xt-1 εt
t=1
(3)
where εt ~ Nd(0, Id) i.i.d. and Xt = Pt =1 εi for t = 1,..., T. This is motivated by the weak
convergence of ZT,d to Zd as T → ∞. Consequently, the proposed procedure relies crucially
on the fact that the first two moments of the asymptotic trace statistic exist and may be
obtained as limits of the corresponding moments of ZT,d.
2 Results
On account of the weak convergence of ZT,d to the asymptotic trace statistic Zd , the first two
moments of Zd exist if the sequence {ZT2,d} is uniformly integrable. A sufficient condition for
this is established in Lemma 2, which states that the fourth moments of ZT,d are uniformly
bounded in T. We start with showing that all moments of ZT,d exist. To ensure that the
inverted matrix appearing in (3) is nonsingular with probability one, we assume T > d.
Lemma 1. Assume that T > d. Then all moments of ZT,d defined by (3) exist.
Proof. As Larsson et al. (2001), we introduce the (T × d) matrices ε
and X = (X1, X2, ..., XT)0 as well as the (T × T) matrices
(ε1, ε2, ..., εT)0
10
and
..... 0 ^
..... 0
0 ■■■ 0
0 10
Then, X = Aε and the (d × d) matrices appearing in (3) can be rewritten as
1T 1
AT := T2 Σ<Xt-1 Xt-1 = T2ε A B BAε,
t=1
1T 1
Bt := T ΣX-1 ^t = TεA'B'ε.
t=1
(4)
Defining D = BA and Y = Dε, we obtain therefore
ZT,d = tr(BT0 AT-1BT) = tr[ε0Dε(ε0D0Dε)-1ε0D0ε]
= tr(ε0PYε) ≤ tr(ε0ε),
(5)
(6)
where PY = Y (Y0Y )-1Y0 denotes the projection matrix onto the column space of Y . The
assumption εt ~ Nd (0 ,Id ) i.i.d. now implies tr( ^ε ) = P T=x ε'tεt ~ χ Td, which completes the
proof, since all moments of a χ2-distributed random variable exist. ¥