regression is rewritten as
zbi+t = Byi(,2t)-1 + vbit , (6)
where z + = y(1l-1 - (biΣ-1 bi)-1 αiΣ-1∆yit, yit = [y(10,yi20]0 and y(1 (y(2 )
are r × 1 (k —r × 1) subvectors of yit. It is interesting to note that z + adopts an
endogeneity correction similar as the estimator of Phillips and Moon (1999).
The important difference is, however, that the latter approach employs a
nonparametric estimate of the endogeneity effect, whereas z + is based on a
parametric endogeneity correction based on a VAR(p) model.
Based on a sequential limit theory, the following theorem states that the
two-step estimator has a normal limiting distribution.
Theorem 1: Let yit be generated as in (1) and Bb2S denotes the least-squares
estimator of B in the regression (6). Furthermore εit and εjt are independent
for i 6= j . If T → ∞ is followed by N → ∞ we have
T √rNvec ( Bb2 S — B) —→ N(0, Ω2 1 ® Σv) ,
where Σv is defined in (4),
1N
ω2 = Jim Tt ∑α2(α0⊥β⊥)-1 αi⊥⊥ςiαi,⊥(β⊥αi,⊥)-1 β⊥2 ,
N→∞ N
i=1
αi,⊥ and β⊥ are orthogonal complements of αi andβ and β⊥,2 is the lower
(n — r) × r block ofβ⊥ .
From this theorem it follows that the long-run parameters are asymptotically
normally distributed and, therefore, the usual tests on the cointegration pa-
rameters involve the usual limiting distributions. In particular, the second-
step regression (6) can be treated as an ordinary regression equation, that
is, the nonstationarity of the regressors and the fact that zbi+t is estimated
can be ignored. Furthermore, it is interesting to note that for finite N and
T → ∞, the estimator are mixed normal, that is, normally distributed with
a stochastic covariance matrix. Therefore, the normal limiting distribution
is expected to yield a reliable approximation even if N is small.