Estimating Long-Run Relationships between Observed ... · Estimating Long-Run Relationships between...
Transcript of Estimating Long-Run Relationships between Observed ... · Estimating Long-Run Relationships between...
FACULTEIT ECONOMIE EN BEDRIJFSKUNDE
TWEEKERKENSTRAAT 2 B-9000 GENT
Tel. : 32 - (0)9 – 264.34.61 Fax. : 32 - (0)9 – 264.35.92
WORKING PAPER
Estimating Long-Run Relationships between Observed Integrated Variables by Unobserved Component Methods
Gerdie Everaert
*
February 2007
2007/452
* SHERPPA, Ghent University, [email protected], http://www.sherppa.be
D/2007/7012/23
Estimating Long-Run Relationships between Observed
Integrated Variables by Unobserved Component Methods ∗
Gerdie Everaert†
February 26, 2007
Abstract
A regression including integrated variables yields spurious results if the residuals containa unit root. Although the obtained estimates are unreliable, this does not automaticallyimply that there is no long-run relation between the included variables as the unit root inthe residuals may be induced by omitted or unobserved integrated variables. This paper usesan unobserved component model to estimate the partial long-run relation between observedintegrated variables. This provides an alternative to standard cointegration analysis. Theproposed methodology is described using a Monte Carlo simulation and applied to investigatepurchasing-power parity.
JEL Classification: C15, C32Keywords: Spurious Regression, Cointegration, Unobserved Component Model, PPP.
1 Introduction
Since the seminal articles of Engle and Granger (1987) and Johansen (1988), cointegration analysis
has become a standard econometric tool for estimating relationships involving integrated variables.
An important drawback of cointegration analysis is that it requires all variables constituting an
equilibrium relation to be included in the analysis. The implication of omitting relevant integrated
variables is double. First, any economic theory indicating a long-run relation between a vector of
integrated variables will fail to yield a cointegrated regression in the presence of omitted integrated
variables. Examples of this problem are legion. Using post-war aggregate U.S. data Rudd and
Whelan (2006), for instance, fail to reject the null hypothesis of no cointegration between con-
sumption, labour income and financial wealth. They argue that a non-stationary component in the
relation between these variables might be induced by non-stationarity of the expected return on
wealth. Second, Engel (2000) shows that standard cointegration tests are seriously biased towards
finding cointegration in the presence of omitted integrated variables. Simulating samples of 100
observations, cointegration tests with a nominal size of 5% are found to have true sizes that range
from 90% to 99% even though there is a non-stationary component that accounts for 42% of the
100-year forecast variance. Combining these two arguments suggests that cointegration analysis is∗I thank Lorenzo Pozzi for useful comments and suggestions. I acknowledge financial support from the Interuni-
versity Attraction Poles Program - Belgian Science Policy, contract no. P5/21.†SHERPPA, Ghent University, [email protected], http://www.sherppa.be
1
not always conclusive in terms of detecting a long-run relation between integrated variables, both
in cases where cointegration is found and in cases where it is rejected.
The interest of this paper is to estimate the long-run relation between integrated variables when
possibly not all of the integrated variables constituting an equilibrium relation are included in the
analysis. As an alternative to standard cointegration analysis, we propose to use the unobserved
component (UC) approach as outlined in, for instance, Harvey (1989) and Durbin and Koopman
(2001). In the UC framework, omitted variables can be treated as unobserved components which
can be inferred from the observed data using the Kalman filter. This allows for estimation of the
long-run relationship between the observed integrated variables using maximum likelihood and for
inference on the long-run parameters using a Wald or a likelihood ratio test even if the variables
are not cointegrated.
Although UC models have recently become very popular to decompose time series into a
number of unobserved components (like trend, cycle, seasonal, ...), they are only rarely used to
filter omitted variables from a long-run relation between integrated variables. Examples are Harvey
et al. (1986), who add an unobserved component to the employment-output relation to account
for the underlying productivity trend, and Sarantis and Stewart (2001) who add an unobserved
component to the consumption-income relationship to account for omitted variables such as wealth.
A major obstruction to its use is that a general UC model is not necessarily identified. Nelson
and Plosser (1982), for instance, show that a difference-stationary process can be decomposed into
a permanent and transitory component in an infinite number of ways depending on the assumed
correlation between innovations to these two unobserved components. Typically identification is
achieved by assuming innovations to be mutually independent. In the context of this paper, this
restriction would imply innovations to the observed and to the unobserved/omitted variables to
be uncorrelated. Obviously, this is a strong restriction. In a number of recent papers (see e.g.
Morley et al., 2003; Morley, 2007) it is shown that an UC model with correlated innovations is
identified provided that it has sufficiently rich dynamics. Therefore, this paper allows for correlated
innovations and checks under which conditions the model is identified.
The performance of the UC framework relative to standard cointegration analysis, both in
terms of estimation and inference, is studied using a Monte Carlo experiment. We consider a
simple dynamic bivariate triangular process similar to the one in Inder (1993). Concerning the
relation between the 2 observed integrated variables, the experiment nests three important cases:
(i) independent random walks, (ii) cointegrated variables and (iii) a long-run relation with an
integrated missing component. In all of these cases, the UC framework provides a consistent and
asymptotically normally distributed estimate for the long-run relation between the 2 observed
variables. In the first and third case, this entails an important improvement over standard estima-
tors, for instance a static ordinary least squares (OLS) estimator, as these yield spurious results in
2
both of these cases. In the second case, the performance of the UC model is similar to that of the
static OLS estimator, which is superconsistent in this case. The nice properties of the UC model
in the third case are of important practical relevance especially as under the various parameter
settings considered in the experiment, standard cointegration tests with a nominal size of 5% have
actual sizes up to 89.1%. This indicates that spurious regressions will wrongly be considered to be
cointegrating regressions in far too many cases. Combining the results for the three cases suggests
that the UC model is a valuable alternative to standard cointegration analysis.
The paper is structured as follows. Section 2 introduces the correlated UC framework as
an alternative to standard single equation cointegration analysis in a simple dynamic bivariate
process. Section 3 presents a Monte Carlo comparison. In section 4, the proposed UC methodology
is applied to testing purchasing-power parity (PPP).
2 Cointegration versus UC analysis
2.1 A simple bivariate process
Consider a dynamic bivariate triangular process
α(L)y1t = µ1 + β(L)y2t + µt, (1)
φ(L)∆y2t = µ2 + ε2t, (2)
µt = νt + ε1t, (3)
νt = νt−1 + η1t, (4)
where y1t and y2t are scalar variables, µ1 and µ2 are constants and the error terms ε1t, ε2t and
η1t are zero mean Gaussian white noise processes with covariance matrix Ω
Ω =
σ2ε1
σε1ε2 σε1η1
σε1ε2 σ2ε2
σε2η1
σε1η1 σε2η1 σ2η1
. (5)
The lag polynomials are defined as α(L) = 1 − α1L − . . . − αpLp, β(L) = β0 + β1L + . . . + βqL
q
and φ(L) = 1 − φ1L + . . . + φrLr. They all have roots outside the unit circle. The choice for
such a simple model is purely for expositional purposes. A highly similar model is used in, among
others, Kremers et al. (1992) and Inder (1993). The main difference is that the error structure of
y1t includes an I(1) component when σ2η16= 0. The non-zero off-diagonal elements of Ω allow for
endogeneity of y2t and collinarity between innovations to νt and innovations to y1t and y2t.
The long-run relation between y2t and y1t can be obtained from rewriting equation (1) as
y1t = λ1 + λ2y2t + γ1(L)∆y1t + γ2(L)∆y2t + ωt, (6)
where λ2 = β(1)/α(1) measures the long-run impact of y2t on y1t, with α(1) = 1− α1 − . . .− αp
3
and β(1) = β0 + β1 + . . .+ βq, and where
λ1 =µ
α(1), γ1(L) =
α(1)− α(L)α(1)(1− L)
, γ2(L) =β(L)− β(1)α(1)(1− L)
, ωt =µt
α(1).
As equation (2) implies that y2t is I(1), y1t and y2t are said to be cointegrated if λ2 6= 0 and
σ2η1
= 0 such that y1t and y2t are I(1) while µt in equation (1) and ωt in equation (6) are I(0). If
λ2 = 0 and/or σ2η16= 0, y1t and y2t are not cointegrated.
2.2 Static cointegration analysis
If λ2 is the primary parameter of interest, Engle and Granger (1987) suggest to use OLS to estimate
a static version of the model in (1)
y1t = λ1 + λ2y2t + υt, (7)
where from using (6) υt = γ1(L)∆y1t + γ2(L)∆y2t + ωt. In estimating (7) the dynamic terms
in ∆y1t and ∆y2t and possible endogeneity of y2t can be ignored due to the superconsistency
property of the OLS estimator when y1t and y2t are cointegrated (Stock, 1987). When y1t and
y2t are not cointegrated, estimating (7) yields spurious results. Testing the null hypothesis that
y1t and y2t are not cointegrated amounts to testing whether υt ∼ I(1) against the alternative
that υt ∼ I(0). Typically this is done using the cointegrating regression Augmented Dickey-Fuller
(CRADF) test. As the CRADF test has low power against stationary alternatives with a root close
to unity, Shin (1994) and Harris and Inder (1994) suggest to test for the null of cointegration by
testing for stationarity of υt building on the Kwiatkowski et al. (1992) unit root test. Leybourne
and McCabe (1994) test the null of cointegration by directly testing whether the variance of the
shocks to the random walk component νt is zero, i.e. H0 : σ2η1
= 0.
2.3 Dynamic cointegration analysis
Although the OLS estimator for λ2 in a static equation like (7) is superconsistent, its asymptotic
distribution depends on nuisance parameters arising from serial correlation in νt and endogeneity
of y2t (Phillips and Durlauf, 1986; Phillips and Hansen, 1990). Moreover, omitting dynamic terms
and ignoring endogeneity leads to substantial small sample biases (Banerjee et al., 1986) and
results in low power of the residual based cointegration tests mentioned above (Kremers et al.,
1992; Zivot, 1994; Banerjee et al., 1996).
Dynamic OLS/GLS estimator
The dynamic OLS (DOLS) estimator suggested by Saikkonen (1991) eliminates nuisance terms
that stem from endogeneity and serial correlation by augmenting (7) with leads and lags of ∆y2t
y1t = λ1 + λ2y2t +k2∑
j=−k1
bj∆y2t−j + εt, (8)
4
where the error term εt is uncorrelated with the regressors at all leads and lags but is in general
serially correlated. Therefore, Stock and Watson (1993) suggest to estimate (8) using feasible
generalised least squares (GLS), referred to as dynamic GLS (DGLS). Monte Carlo evidence
in Stock and Watson (1993) shows that the DOLS and DGLS estimators yield an important
improvement over the static OLS estimator both in terms of estimation and inference.
ECM estimator
Banerjee et al. (1986) suggest to estimate λ2 by estimating the dynamic equation (1), which is
conveniently written in error correction model (ECM) form as
δ1(L)∆y1t = δ′2(L)∆y2t − α(1) (y1t−1 − λ1 − λ′2y2t−1) + µt, (9)
where
δ1(L) =α(L)− α(1)L
(1− L), δ2(L) =
β(L)− β(1)L(1− L)
.
Using Monte Carlo simulations Inder (1993) demonstrates that in a dynamic setting the ECM
estimator provides precise estimates and valid inference even in the presence of endogenous vari-
ables. To test for the null of no cointegration Kremers et al. (1992), Zivot (1994) and Banerjee
et al. (1996) suggest to use a t-test for α(1) = 0 . Using Monte Carlo simulations, this ECM test
for cointegration is found to be more powerful than residual based tests.
2.4 An unobserved component framework
The results of the above mentioned estimation and test procedures are not always conclusive in
terms of detecting the long-run relation between y1t and y2t. Upon finding no cointegration one
cannot automatically conclude that there is no long-run relation between y1t and y2t, as the unit
root in µt may be induced by omitted or unobserved I(1) variables. In this case νt represents (a
linear combination of) omitted or unobserved I(1) variables which ought to be included in (1) for
µt to be I(0). Although the presence of a unit root in µt implies that the results obtained from
estimating equation (7), (8) or (9) using OLS/GLS are spurious, the model in equations (1)-(4)
can be cast into a linear Gaussian state-space (SS) representation and estimated using maximum
likelihood (ML). As such, it is possible to obtain a non-spurious estimate for the long-run relation
between y1t and y2t even if they are not cointegrated.
State-space representations
The general SS form1 is given by
yt = Axt + Zαt + εt, εt ∼ N(0,H), t = 1, . . . , T, (10)
αt+1 = Sαt +Rηt, ηt ∼ N(0, Q), E [ε′tηt] = G, (11)1See e.g. Harvey (1989) or Durbin and Koopman (2001) for an extensive overview of SS models.
5
where yt is a p×1 vector of p observed endogenous variables modelled in the observation equation
(10), xt is a k × 1 vector of k observed exogenous or predetermined variables and αt is a m × 1
vector of m unobserved states modelled in the state equation (11). The matrices A,Z, S,R,H,Q
and G are time-invariant but generally depend on an unknown parameter vector ψ.
If the explanatory variable y2t is weakly exogenous in equation (1), the SS representation of
the UC model in equations (1)-(5) is given by the observation equation
y1t =[
(1−α(L))L β(L)
] [y1t−1
y2t
]+ νt + ε1t, (12)
with the state equation being equation (4) and where H = σ2ε1, Q = σ2
η1and G = σε1η1 . Without
loss of generality, the constant µ1 is included in the state variable νt. The condition that y2t is
weakly exogenous means that no relevant information to the estimation of the unknown parameters
in equation (1) is lost by conditioning on y2t (Engle et al., 1983), i.e. it is not necessary to construct
a model for y2t. This is the case if σε1ε2 = 0 and σε2η1 = 0.
If the explanatory variable y2t is not weakly exogenous, the reduced form of the UC model in
equations (1)-(2) is given by
y1t = µ1 + β0µ2 + α′(L)y1t−1 + β′(L)y2t−1 + νt + ε1t + β0ε2t, (13)
y2t = µ2 + φ′(L)y2t−1 + ε2t, (14)
where
α′(L) =1− α(L)
L; β′(L) =
β(L)− β0(1− φ′(L)L)L
; φ′(L) =1 + (L− 1)φ(L)
L.
This reduced form UC model can be cast in SS form with observation equation
[y1t
y2t
]=
[α′(L) β′(L)
0 φ′(L)
] [y1t−1
y2t−1
]+
[β0 1 β0
1 0 1
] ε2t
νt
µ2
+[ε1t
0
], (15)
and state equation ε2t
νt
µ2
=
0 0 00 1 00 0 1
ε2t−1
νt−1
µ2
+
1 00 10 0
[ε2t
η1t
], (16)
where H = σ2ε1, Q =
[σ2
ε2σε2η1
σε2η1 σ2η1
]and G =
[σε1ε2 σε1η1
]′ . Without loss of generality, the
constant µ1 is again included in the state variable νt.
Identification
It is not immediately obvious that the SS models presented above are identified. First consider
the single equation model in equations (12) and (4). In order to check identification, first derive
6
the stationary UC autoregressive moving-average (ARMA) representation of the model which is,
under the assumption that y2t is weakly exogenous, given by
α(L)∆y1t − β(L)∆y2t = η1t + ∆ε1t. (17)
The right-hand side of equation (17) will have nonzero autocovariances through lag 1. Then, by
Granger’s lemma (Granger and Newbold, 1986) the reduced-form ARMA representation of the
model is given by
α(L)∆y1t − β(L)∆y2t = ξ(L)et ≡ ςt, (18)
where the MA lag polynomial ξ(L) is of order 1. After normalising ξ0 = 1, ξ1 and σ2e , are obtained
such that the autocovariances of the right-hand side of (18) match with those of (17). The SS
model in (17) is identified if its parameters can be inferred from the reduced form in (18). As the
left-hand side of both equations is equal, the (p+q+1) parameters in the AR lag polynomials α(L)
and β(L) are always identified. The remaining 3 structural coefficients (σ2ε1, σ2
η1, σε1η1) cannot be
inferred from the 2 coefficients ξ1 and σ2e on the right-hand side of (18), though. The relation
between the reduced form and the UC parameters is given by
γ0 = 2(σ2ε1
+ σε1η1) + σ2η1,
γ1 = −(σ2ε1
+ σε1η1),
γτ = 0 for τ > 2,
where the values for the autocovariances γτ = cov (ςt, ςt−τ ) for τ = 0, 1 can be estimated from
the data through the reduced-form ARMA representation of the model. This shows that σ2ε1
and
σε1η1 are not individually identified but only their sum is. Morley et al. (2003) point out that a
separate identification of σ2ε1
and σε1η1 requires ε1t to exhibit rich enough dynamics. In stead of
generalising equation (3) by allowing ε1t to be generated from a more general ARMA process we
impose σε1η1 = 0. This identifying restriction affects the estimate for σ2ε1
but leaves the estimates
for the other structural coefficients unaffected. As we are only interested in estimating the long-run
relation between y1t and y2t, this restriction is without loss of generality.
Next, consider the stationary UC vector autoregressive moving-average (VARMA) representa-
tion of the multiple equation framework in(15)-(16)
φ(L)α(L)∆y1t − φ(L)(β(L)− β0)∆y2t = β0µ2 + φ(L)∆ε1t + β0ε2t + φ(L)η1t, (19)
φ(L)∆y2t = µ2 + ε2t. (20)
The right-hand side of (19) will have nonzero autocovariances through lag (r+1). Then, by
Granger’s lemma (Granger and Newbold, 1986) the reduced-form VARMA representation of the
7
model is given by
φ(L)α(L)∆y1t − φ(L)(β(L)− β0)∆y2t = ω1 + ξ11(L)e1t + ξ12(L)e2t ≡ ς1t, (21)
φ(L)∆y2t = ω2 + e2t ≡ ς2t, (22)
where the MA lag polynomials ξ11(L) and ξ12(L) are both of order (r + 1) with normalisations
ξ11,0 = 1 and ξ12,0 = 0. The remaining coefficients in ξ11(L) and ξ12(L) and σ2e1, σ2
e2, σe1e2 are
obtained such that the autocovariances of the right-hand sides of (21)-(22) match with those of
(19)-(20). The UC model is identified if its parameters can be inferred from the reduced form.
First note that the left-hand sides of (19)-(20) and (21)-(22) are equal. This shows that the
(p+q+r) parameters in the AR lag polynomials α(L), (β(L)−β0) and φ(L) are always identified.
Second, the coefficients µ2 and β0 on the right-hand side of (19)-(20) and the 6 variance-covariances
in Ω need to be inferred from the 2(r + 1) MA coefficients, the two intercept terms ω1 and ω2
on the right-hand side of (21)-(22) and the 3 variance-covariance parameters σ2e1, σ2
e2, σe1e2 . A
necessary condition for identification is that r > 1, i.e. the reduced form contains as least as much
parameters as the structural UC model. In order to check whether this necessary condition is also
sufficient for identification, a one-to-one mapping between the reduced-form and the structural
parameters should be established, though. Consider for instance a model with µ2 6= 0 and r = 1.
The parameters in the AR part of this model can be estimated directly from the data while the
observable moments of the MA part of the model are the intercept terms ω1 and ω2 and the
autocovariances
γ0,11 = 2(1 + φ1 + φ21)(σ
2ε1
+ σε1η1) + β20σ
2ε2
+ (1 + φ21)σ
2η1
+ 2β0σε1ε2 + 2β0σε2η1 ,
γ0,12 = γ0,21 = β0σ2ε2
+ σε1ε2 + σε2η1 ,
γ0,22 = σ2ε2,
γ1,11 = −(1 + φ1)2(σ2ε1
+ σε1η1)− φ1σ2η1− β0(1 + φ1)σε1ε2 − β0φ1σε2η1 ,
γ1,12 = β1σ2ε2− (1 + φ1)σε1ε2 − φ1σε2η1 , (23)
γ2,11 = φ1(σ2ε1
+ σε1η1) + β0φ1σε1ε2 ,
γ2,12 = φ1σε1ε2 ,
γτ,2j = 0 for τ > 1 and γτ,1j = 0 for τ > 2,
where γτ,ij = cov (ςit, ςjt−τ ) with i, j = 1, 2. The 7 non-zero autocovariances together with the
intercepts terms ω1 and ω2 provide 9 pieces of information from which in principle the remaining
8 UC coefficients (µ2, β0, σ2ε1, σ2
ε2, σ2
η1, σε1ε2 , σε1η1 and σε2η1) can be inferred. Inspection of the
autocovariances reveals that σ2ε1
and σε1η1 are again not individually identified, though. Without
loss of generality the model is again identified by setting σε1η1 = 0. Imposing this identifying
restriction implies that the remaining structural coefficients are identified either if µ2 6= 0 or if
8
r > 1. In the former, the non-zero drift term µ2 implies that both µ2 and β0 can be identified
from the intercept terms ω1 and ω2 while, even if r = 0, the remaining 5 parameters in Ω can
be identified from the 5 non-zero autocovariances in (23). In the latter, the zero drift term µ2
implies that β0 should also be identified from the autocovariances in (23). This requires at least
6 non-zero autocovariances, i.e. r > 1.
Intuitively, identification of the long-run impact of y2t on y1t requires (i) y2t to be weakly
exogenous such that it is orthogonal to the unobserved component νt or (ii) the dynamic behaviour
of y2t to differ from that of the unobserved component νt once innovations to these components
are allowed to be cross-correlated.
ML estimation and inference
Provided that the UC model is identified, the likelihood for the linear Gaussian SS model in (10)-
(11) can be calculated by a routine application of the Kalman filter and maximised with respect
to the unknown parameter vector ψ using an iterative numerical procedure. Pagan (1980) shows
that the resulting ML estimator ψ is consistent and asymptotically normally distributed provided
that (i) ψ is an interior point of the parameter space and (ii) the transition matrix S does not
contain unknown polynomials with roots inside the unit circle. As λ2 is an interior point and S
does not contain unknown parameters, inference on the long-run relation between y1t and y2t is
possible using a standard Wald or likelihood ratio (LR) test. Note that testing whether y1t and
y2t are cointegrated implies testing whether σ2η1
= 0, which is on the boundary of the parameter
space. In this case, the distribution of a Wald or LR test is non-standard but can in principle be
obtained using Monte Carlo simulation. As the interest of this paper is to estimate the long-run
relationship between y1t and y2t irrespectively of whether they are cointegrated or not, we don’t
want to test this hypothesis, though.
3 Monte Carlo experiment
In this section, the performance of 5 alternative estimators for the long-run relationship between
integrated variables is compared using a Monte Carlo simulation.
3.1 Design
As before, we consider the model in (1)-(5), setting p = q = r = 1. Data are generated under a
variety of settings for the parameter vector ψ =(α1, β0, λ2, φ1, µ2, σ
2ε1, σ2
ε2, σ2
η1, σε1ε2 , σε1η1 , σε2η1
).
For each parameter setting we generate 2500 series of length T . First, we draw a sample of T
observations for ε1t, ε2t and η1t out of a multivariate normal distribution with variance-covariance
matrix Ω. Next, we generate y2t and νt by first drawing initial values for ∆y2t and ∆νt from their
stationary distribution and then calculating initial values for y2t and νt from a diffuse initialisation
9
on the level
y2,1 = y2,0 + ∆y2,1, where ∆y2,1 = (1− φ21)−1/2ε2,1 and y2,0 ∼ N(0,∞), (24)
ν1 = ν0 + ∆ν1, where ∆ν1 = η1 and ν0 ∼ N(0,∞). (25)
Data for t = 2, . . . , T are then generated using (2) and (4) respectively. In order to generate data
for y1t, first define the latent variable ζt to be the deviation of y1t from its equilibrium level implied
by the levels of µ1, y2t and νt
ζt = y1t −1
(1− α1)(µ1 + (β0 + β1)y2t + νt) . (26)
Using (1), ζt can be written as
ζt = α1ζt−1 + ξt, (27)
where ξt = ε1t −α1β0 + β1
(1− α1)∆y2t −
α1
(1− α1)η1t.
Setting ∆y2t = (1− φ21)−1/2ε2,1 from (24), ζt can be initialised from (27) as
ζ1 = ξ1(1− α21)−1/2. (28)
After generating ζt for t = 2, . . . , T using (27) and given data on y2t and νt, y1t can then be
calculated from (26).
We consider five alternative estimators: (i) the static OLS estimator in equation (6), (ii) the
DGLS estimator in equation (8), (iii) the ECM estimator in equation (9), (iv) the univariate UC
model (UC U) in equations (12) and (4) and (v) the multivariate UC model (UC M) in equations
(15)-(16). The DGLS estimator is implemented by setting as a rule of thumb k1 = k2 = int[T 1/3
](see Saikkonen, 1991). Selecting an alternative number of leads and lags did not have a noteworthy
impact on the main results. The ECM, UC U and UC M estimators are implemented by setting,
if relevant, p = q = r = 1.
The design of the Monte Carlo experiment nests three special cases. First (Case I), when
λ2 = 0 and σ2η16= 0, there is no long-run relationship between the random walk processes y1t
and y2t. Second (Case II), when λ2 6= 0 and σ2η1
= 0, y1t and y2t are cointegrated. Third (case
III), when λ2 6= 0 and σ2η16= 0, there is a long-run relationship between y1t and y2t but they
are not cointegrated. For each of these cases we analyse the performance of the 5 alternative
estimators in terms of estimation and inference. Next to the median bias and the root mean
squared error (Rmse), we report the percentiles of the empirical distribution of λ2.2 We also check
the performance of standard cointegration tests, i.e. a CRADF test on the residuals from the OLS2We report the median bias instead of the mean bias and percentiles along with the Rmse as the dynamic
estimators ECM, UC U and UC M for λ2 = (β0 + β1)/(1− α1) do not have finite moments, especially when α1 isclose to 1. It should be noted that the percentiles are not necessarily finite either but they should be less vulnerableto large outliers in the distribution of λ2.
10
and DGLS regressions and a t-test on the error-correction term in the ECM regression. Critical
values for these cointegration test statistics are obtained by simulating their distributions under the
null hypothesis of no cointegration, i.e by first drawing y1t and y2t from a standardised multivariate
random walk as ∆yt ∼ IN (0, Ik) and next calculating the considered cointegration test statistics.
The obtained critical values for the OLS and the ECM-based cointegration tests coincide with
those reported in Mackinnon (1996) and Ericsson and MacKinnon (2002) respectively. The reason
why critical values are simulated is that no such values are available for a cointegration test based
on DGLS residuals.
3.2 Results
Results for Case I are reported in Table 1. It is a well-known fact that the OLS estimator yields
spurious results in this case. This can be seen from the fact that, although the median bias is
negligible, the null hypothesis that λ2 = 0 is rejected (with a nominal size of 5%) in 74.9% of the
cases for T = 50. This problem even aggravates as T grows large. The reason for this huge size
bias is that the OLS estimator does not converge in probability and the t- statistic does not have
a well-defined asymptotic distribution. Note that the DGLS and the ECM estimators face the
same problem, although somewhat less pronounced. The UC U and UC M estimators are both
unbiased and more or less correctly sized. This shows that these estimators are not vulnerable to
the spurious regression problem. In terms of estimation, the UC U estimator is much more precise
than the UC M estimator, which is overparameterised in this case.
Results for the case where y1t and y2t are cointegrated (Case II) are reported in Table 2. The
performance of the estimators is explored both in a static setting (see Panels (a) and (b) where
α1 = 0 and β0 = λ2 = 1) and in a dynamic setting (see Panels (c) and (d) where α1 = 0.5,
β0 = 0.2). In both settings, we discriminate between a case where y2t is exogenous (σε1ε2 = 0)
and a case where y2t is endogenous (σε1ε2 = 0.5). In terms of estimation, the OLS estimator
is superconsistent, i.e. it converges to the true population value at a rate faster than in normal
asymptotics even if dynamic terms and endogeneity issues are ignored. The results in Panels (b),
(c) and (d) show that even in a small sample (T = 50) the bias induced by endogeneity and
ignoring dynamic terms is small and, as implied by the superconsistency property, disappears
quickly as T increases. In terms of inference, the OLS estimator has a correct size in the static
setting without endogeneity. The effects of endogeneity on the distribution of the OLS estimator
are minimal, while in the dynamic setting the size is unacceptably high. Turning to the DGLS
and ECM estimators, only the latter yields an improvement on the OLS estimator in terms of
estimation. Despite its overparameterisation in the static setting, the ECM estimator has a bias
and Rmse which are highly similar to those of the OLS estimator. In the dynamic setting, the
ECM estimator has a notable lower bias and Rmse. In terms of inference, both the DGSL and
11
Table 1: Monte Carlo results case I: λ2 = 0, σ2η1 = 1
(α1 = 0, β0 = 0, φ1 = 0.5, µ2 = 0.25, σ2ε1
= 0, σ2ε2
= 1, σε1ε2 = 0, σε1η1 = 0, σε2η1 = 0)T Estimator bias(λ2) Rmse (λ2) p2,5% p97,5% P (λ2) P (CI)
50 OLS −0.001 0.278 −0.581 0.568 0.749 0.040DGLS 0.005 0.362 −0.750 0.769 0.381 0.076ECM 0.001 > 10 −1.307 1.518 0.291 0.062UC U −0.003 0.138 −0.281 0.271 0.062 -UC M −0.003 0.346 −0.615 0.548 0.045 -
100 OLS −0.006 0.216 −0.436 0.389 0.858 0.048DGLS −0.001 0.219 −0.443 0.418 0.395 0.084ECM −0.004 4.531 −0.815 0.741 0.359 0.071UC U 0.002 0.095 −0.187 0.186 0.062 -UC M −0.004 0.263 −0.404 0.385 0.045 -
250 OLS −0.002 0.143 −0.273 0.290 0.911 0.050DGLS 0.004 0.129 −0.246 0.266 0.454 0.098ECM 0.001 1.657 −0.399 0.437 0.418 0.068UC U 0.002 0.058 −0.113 0.115 0.055 -UC M 0.002 0.125 −0.246 0.244 0.054 -
500 OLS −0.002 0.101 −0.199 0.199 0.941 0.062DGLS −0.001 0.095 −0.195 0.190 0.504 0.120ECM −0.003 2.755 −0.252 0.274 0.434 0.078UC U 0.000 0.041 −0.078 0.082 0.052 -UC M −0.001 0.090 −0.181 0.176 0.076 -
Notes: bias(bλ2) and Rmse(bλ2) are the Monte Carlo median bias and root mean squared error of bλ2,respectively. p2,5% and p97,5% are the 2,5% and 97,5% percentiles, respectively, of the Monte
Carlo distribution of bλ2. P (bλ2) is the rejection rate at the 5% level of the test statistic testingbλ2 = λ2. For the OLS, DGLS and ECM estimators this is a standard t-test. For the UC Uand UC M this is a Wald test. P (CI) is the rejection rate at the 5% level for a test for thenull hypothesis of no cointegration. For the OLS and DGLS estimators, this is a DF test on theestimated residuals. For the ECM estimator this is a t-test on the error-correction term. Thesimulated 5% critical values for the OLS, DGLS and ECM estimators respectively are (i) -3.47,-4.34 and -3.30 for T = 50, (ii) -3.40, -3.64 and -3.25 for T = 100, (iii) -3.36, -3.32 and -3.24 forT = 250 and (iv) -3.34, -3.24 and -3.23 for T = 500.
ECM estimator have a more or less correct size in all cases. Only in the case of endogeneity, the
size of the ECM estimator is somewhat different from 5%. Consistent with the finding of Inder
(1993), the relative performance of the ECM estimator suggests that it is better to overspecify
the dynamics of the model than to underspecify. Interestingly, the performance of the UC U and
the UC M estimators are almost identical to that of the ECM estimator. This shows that adding
an unobserved I(1) component to the error structure of the ECM does not impair performance in
the special case of cointegration between y1y and y2t.
Results for the case where y1t and y2t have a long-run relationship but are not cointegrated
(Case III) are reported in Table 3. The OLS, DGLS and ECM estimators all yield spurious results.
As in Case I, the median bias is negligible but the size of testing the hypothesis that λ2 = λ2 = 1
is unacceptably high. Again, the problem aggravates as T grows large. Important to note is
that cointegration tests also have a large size bias, especially in the static case and for the ECM
12
estimator in the dynamic case. This is consistent with the conclusion of Engel (2000) that standard
cointegration tests are strongly biased towards rejecting the null hypothesis of no cointegration in
the presence of an omitted permanent component. The UC U estimator yields an improvement
both in terms of estimation precision and inference. However, its size only converges slowly to 5%
as T grows large, especially for σε1ε2 6= 0. The UC M estimator is less precise compared to the
UC U estimator, i.e. its Rmse is about the same as for the OLS, DGLS and ECM estimators, but
has better size properties compared to the UC U estimator. Note that in a small sample the size
is also too large, though.
Tables 4-5 further explore the performance of the UC U and UC M estimators in Case III for a
number of alternative values for the parameters α1, β0, φ1, µ2 and σ2η1
. Table 4 reports results for
a setting where innovations to the dependent variable y1t are correlated with innovations to the
unobserved component νt (i.e. σε1η1 = 0.5). Table 5 reports results for a setting where innovations
to the explanatory variable y2t are correlated with innovations to the unobserved component νt
(i.e. σε2η1 = 0.5). As the OLS, DGLS and ECM estimators all yield spurious results for σ2η16= 0,
comparable to the results in Table 3, they were excluded from the tables to economise on space
(results available on request). Comparing Panel (b) from Table 4 with Panel (c) from Table 3
shows that, although σε1η1 is set to zero as an identifying restriction for both the UC U and
UC M estimator, the performance of both estimators is not affected, as indicated in section 2.4,
by simulating data with σε1η1 = 0.5. The results in Table 4 further show that the estimation
precision deteriorates as (i) α1 increases (compare Panels (a), (b) and (c)), (ii) the dynamics in
the explanatory variable y2t are less rich (see Panel (e) for less inertia in y2t, i.e. lower value of φ1,
and Panel (f) for zero drift, i.e. µ2 = 0) and (iii) as the variance of the unobserved component νt
increases (compare Panels (b), (g) and (h)). Given its overparameterisation, UC M is again less
precise than UC U. Inference is similar over all parameter settings included in Table 4. Turning
to Table 5, it is immediately clear that the UC U estimator performs poorly, both in terms of
estimation and inference when σε2η1 6= 0. The performance of the UC M estimator deteriorates a
little bit, especially for small T , but is still satisfactory for larger T .
Summarising, the Monte Carlo simulation shows that standard estimators and cointegration
tests can only deal with cases where y1t and y2t are either independent random walks or cointe-
grated. In the case of a long-run relation with an integrated missing component, these estimators
yield spurious results with standard cointegration tests indicating these results to be a cointegra-
tion regression in far too many cases. The UC approach yields results similar to those of standard
estimators in the case of cointegration and does not yield spurious results in the cases where y1t
and y2t are independent random walks or related in the long-run with an unobserved component.
This suggests that it is better to overspecify than to underspecify the error structure of the model.
13
Table
2:
Monte
Carl
ore
sult
sfo
rca
seII
:λ
2=
1,σ
2 η1
=0
(φ1
=0.
5,µ
2=
0.25
,σ
2 ε1
=1,σ
2 ε2
=1,σ
ε1η1
=0,σ
ε2η1
=0)
TE
stim
ator
bias
(λ2)
Rm
se(λ
2)
p2,5
%p97,5
%P
(λ2)
P(CI)
bias
(λ2)
Rm
se(λ
2)
p2,5
%p97,5
%P
(λ2)
P(CI)
(a)α
1=
0,β
0=
1,σ
ε1ε2
=0
(b)α
1=
0,β
0=
1,σ
ε1ε2
=0.
5
50O
LS
−0.
001
0.02
50.
946
1.04
90.
049
1.00
0−
0.00
10.
023
0.94
81.
046
0.03
91.
000
DG
LS
−0.
001
0.05
30.
894
1.09
70.
067
0.99
6−
0.00
10.
043
0.90
81.
081
0.07
40.
994
EC
M−
0.00
10.
026
0.94
41.
053
0.06
41.
000
−0.
005
0.02
50.
934
1.03
30.
090
1.00
0U
CU
−0.
001
0.02
70.
941
1.05
30.
094
-−
0.00
50.
029
0.92
11.
033
0.11
2-
UC
M−
0.00
10.
029
0.93
71.
055
0.06
0-
−0.
001
0.02
50.
948
1.04
90.
064
-
100
OLS
0.00
00.
009
0.98
31.
017
0.04
21.
000
0.00
00.
008
0.98
41.
017
0.03
91.
000
DG
LS
0.00
00.
014
0.97
51.
029
0.05
71.
000
0.00
00.
012
0.98
01.
023
0.05
41.
000
EC
M0.
000
0.00
90.
982
1.01
80.
052
1.00
0−
0.00
10.
008
0.98
01.
012
0.08
61.
000
UC
U0.
000
0.01
00.
982
1.01
80.
066
-−
0.00
10.
011
0.97
21.
012
0.09
8-
UC
M0.
000
0.01
00.
981
1.01
90.
042
-0.
000
0.00
80.
984
1.01
60.
048
-
250
OLS
0.00
00.
002
0.99
61.
004
0.04
21.
000
0.00
00.
002
0.99
61.
004
0.04
31.
000
DG
LS
0.00
00.
002
0.99
61.
004
0.04
31.
000
0.00
00.
002
0.99
61.
004
0.04
41.
000
EC
M0.
000
0.00
20.
996
1.00
40.
045
1.00
00.
000
0.00
20.
996
1.00
30.
061
1.00
0U
CU
0.00
00.
002
0.99
61.
004
0.05
2-
0.00
00.
003
0.99
41.
003
0.07
2-
UC
M0.
000
0.00
20.
996
1.00
40.
031
-0.
000
0.00
20.
996
1.00
40.
028
-
(c)α
1=
0.5,β
0=
0.2,σ
ε1ε2
=0
(d)α
1=
0.5,β
0=
0.2,σ
ε1ε2
=0.
5
50O
LS
−0.
036
0.09
30.
761
1.08
00.
412
0.46
3−
0.02
80.
075
0.79
71.
056
0.41
00.
516
DG
LS
−0.
004
0.09
70.
803
1.17
70.
110
0.49
0−
0.00
30.
079
0.82
31.
146
0.12
20.
496
EC
M−
0.00
40.
054
0.87
31.
099
0.07
50.
998
−0.
008
0.05
20.
862
1.07
00.
060
0.94
2U
CU
−0.
004
0.05
50.
870
1.09
80.
100
-−
0.00
90.
057
0.84
51.
070
0.08
5-
UC
M−
0.00
20.
059
0.86
51.
110
0.07
5-
−0.
003
0.04
90.
888
1.09
60.
076
-
100
OLS
−0.
010
0.03
30.
910
1.03
50.
364
0.98
8−
0.00
80.
026
0.93
01.
025
0.37
00.
992
DG
LS
0.00
00.
026
0.95
11.
053
0.07
60.
999
−0.
002
0.02
30.
962
1.04
40.
078
0.99
8E
CM
0.00
00.
018
0.96
31.
034
0.05
61.
000
−0.
002
0.01
80.
962
1.02
60.
037
1.00
0U
CU
0.00
00.
020
0.96
21.
037
0.07
9-
−0.
002
0.01
80.
961
1.02
60.
061
-U
CM
0.00
00.
023
0.96
11.
037
0.05
0-
0.00
00.
017
0.96
71.
032
0.04
8-
250
OLS
−0.
002
0.00
70.
983
1.01
10.
342
1.00
0−
0.00
20.
006
0.98
61.
008
0.36
81.
000
DG
LS
0.00
00.
004
0.99
21.
008
0.04
81.
000
0.00
00.
004
0.99
31.
008
0.05
81.
000
EC
M0.
000
0.00
40.
993
1.00
70.
049
1.00
00.
000
0.00
40.
993
1.00
60.
019
1.00
0U
CU
0.00
00.
004
0.99
31.
007
0.06
4-
0.00
00.
004
0.99
21.
006
0.02
9-
UC
M0.
000
0.00
40.
992
1.00
80.
037
-0.
000
0.00
40.
993
1.00
70.
027
-
Note
s:See
Table
1.
14
Table
3:
Monte
Carl
ore
sult
sfo
rca
seII
I:λ
2=
1,σ
2 η1
=0.5
(φ1
=0.
5,µ
2=
0.25
,σ
2 ε1
=1,σ
2 ε2
=1,σ
ε1η1
=0,σ
ε2η1
=0)
TE
stim
ator
bias
(λ2)
Rm
se(λ
2)
p2,5
%p97,5
%P
(λ2)
P(CI)
bias
(λ2)
Rm
se(λ
2)
p2,5
%p97,5
%P
(λ2)
P(CI)
(a)α
1=
0,β
0=
1,σ
ε1ε2
=0
(b)α
1=
0,β
0=
1,σ
ε1ε2
=0.
5
50O
LS
−0.
002
0.19
80.
591
1.40
50.
692
0.67
6−
0.00
20.
200
0.58
71.
400
0.70
00.
670
DG
LS
0.00
30.
267
0.46
81.
563
0.45
20.
498
0.00
80.
268
0.44
91.
538
0.44
70.
423
EC
M−
0.00
10.
225
0.52
61.
416
0.45
30.
702
−0.
015
0.23
00.
491
1.37
00.
486
0.75
6U
CU
−0.
005
0.12
90.
746
1.26
20.
171
-−
0.04
30.
126
0.70
51.
173
0.13
7-
UC
M−
0.00
40.
242
0.56
01.
431
0.18
2-
0.00
10.
228
0.58
01.
418
0.16
5-
100
OLS
−0.
002
0.15
20.
689
1.27
40.
838
0.73
0−
0.00
20.
154
0.67
81.
291
0.84
20.
734
DG
LS
−0.
003
0.16
60.
656
1.30
90.
576
0.69
60.
000
0.16
40.
655
1.31
60.
559
0.59
3E
CM
−0.
004
0.15
80.
665
1.29
20.
586
0.77
5−
0.01
10.
163
0.64
31.
288
0.61
30.
816
UC
U−
0.00
10.
082
0.83
11.
168
0.10
7-
−0.
044
0.08
90.
792
1.09
90.
104
-U
CM
−0.
003
0.16
30.
707
1.28
10.
092
-−
0.00
30.
160
0.69
91.
283
0.08
4-
250
OLS
−0.
001
0.10
10.
807
1.20
60.
906
0.84
5−
0.00
10.
102
0.79
31.
215
0.90
90.
840
DG
LS
−0.
001
0.10
10.
803
1.20
50.
666
0.84
0−
0.00
10.
103
0.79
81.
221
0.63
90.
739
EC
M−
0.00
20.
103
0.80
71.
212
0.67
10.
861
−0.
004
0.10
40.
792
1.20
50.
683
0.89
1U
CU
0.00
10.
049
0.90
41.
094
0.06
2-
−0.
043
0.06
40.
857
1.04
70.
137
-U
CM
0.00
10.
089
0.82
61.
176
0.04
8-
0.00
10.
090
0.81
51.
180
0.04
2-
(c)α
1=
0.5,β
0=
0.2,σ
ε1ε2
=0
(d)α
1=
0.5,β
0=
0.2,σ
ε1ε2
=0.
5
50O
LS
−0.
038
0.39
90.
089
1.74
40.
739
0.04
3−
0.03
80.
401
0.10
71.
756
0.74
90.
028
DG
LS
−0.
005
0.50
7−
0.09
12.
023
0.37
20.
074
−0.
003
0.49
0−
0.05
31.
962
0.36
90.
080
EC
M0.
076
2.24
9−
0.02
13.
128
0.38
90.
330
0.05
67.
607
−0.
876
3.33
40.
297
0.08
4U
CU
0.00
70.
353
0.49
31.
739
0.19
2-
−0.
041
>10
0.34
91.
964
0.27
9-
UC
M0.
012
0.49
90.
062
1.98
90.
182
-0.
015
0.91
80.
135
2.00
60.
167
-
100
OLS
−0.
015
0.30
70.
347
1.53
30.
857
0.06
8−
0.01
30.
308
0.34
41.
558
0.85
50.
028
DG
LS
−0.
007
0.30
50.
342
1.58
20.
394
0.08
6−
0.00
70.
300
0.35
91.
578
0.38
80.
087
EC
M0.
042
1.51
80.
289
2.03
30.
484
0.43
20.
032
6.68
1−
0.17
12.
315
0.36
80.
130
UC
U0.
006
0.19
60.
636
1.41
80.
124
-−
0.00
3>
100.
530
1.53
50.
234
-U
CM
0.00
1>
100.
409
1.59
90.
122
-0.
002
0.32
00.
372
1.62
10.
098
-
250
OLS
−0.
004
0.20
30.
604
1.40
20.
913
0.11
1−
0.00
40.
204
0.58
71.
425
0.91
50.
034
DG
LS
0.00
30.
182
0.65
51.
369
0.44
90.
098
0.00
40.
182
0.65
61.
368
0.43
20.
084
EC
M0.
018
0.35
10.
542
1.55
00.
571
0.51
60.
016
2.89
80.
423
1.67
70.
437
0.16
2U
CU
0.00
30.
111
0.77
81.
224
0.07
5-
0.02
20.
188
0.74
01.
306
0.12
5-
UC
M0.
005
0.18
10.
654
1.36
10.
054
-0.
004
0.18
20.
630
1.36
60.
050
-
Note
s:See
Table
1.
15
Table
4:
Monte
Carl
ore
sult
sfo
rca
seII
I:λ
2=
1,σ
2 ε1
=1,σ
2 ε2
=1,σ
ε1ε2
=0,σ
ε1η1
=0.5
,σ
ε2η1
=0
TE
stim
ator
bias
(λ2)
Rm
se(λ
2)
p2,5
%p97,5
%P
(λ2)
bias
(λ2)
Rm
se(λ
2)
p2,5
%p97,5
%P
(λ2)
(a)α
1=
0.25
,β
0=
0.2,φ
1=
0.5,µ
2=
0.25
,σ
2 η1
=0.
5(b
)α
1=
0.5,β
0=
0.2,φ
1=
0.5,µ
2=
0.25
,σ
2 η1
=0.
5
50U
CU
−0.
019
0.19
80.
618
1.41
30.
196
−0.
010
0.41
20.
449
1.73
00.
219
UC
M−
0.00
10.
323
0.38
41.
612
0.21
20.
011
0.49
50.
021
1.97
50.
203
100
UC
U0.
001
0.12
00.
765
1.24
20.
117
0.00
20.
203
0.62
91.
429
0.13
7U
CM
−0.
002
0.19
60.
601
1.37
50.
129
0.00
70.
322
0.37
21.
614
0.13
9
250
UC
U0.
000
0.07
20.
854
1.14
40.
073
0.00
00.
116
0.76
81.
239
0.07
9U
CM
0.00
40.
123
0.74
71.
229
0.06
50.
010
0.18
70.
621
1.34
60.
071
(c)α
1=
0.75
,β
0=
0.2,φ
1=
0.5,µ
2=
0.25
,σ
2 η1
=0.
5(d
)α
1=
0.5,β
0=
0,φ
1=
0.5,µ
2=
0.25
,σ
2 η1
=0.
5
50U
CU
−0.
049
>10
−0.
111
3.13
20.
213
−0.
008
0.32
50.
436
1.70
50.
218
UC
M0.
027
1.58
3−
1.13
33.
497
0.17
10.
013
0.49
70.
061
1.99
40.
205
100
UC
U−
0.01
81.
527
0.28
21.
937
0.13
70.
003
0.19
80.
633
1.40
30.
124
UC
M0.
013
0.87
9−
0.23
12.
311
0.08
10.
007
0.32
00.
391
1.61
70.
128
250
UC
U0.
000
0.23
20.
560
1.47
70.
080
0.00
00.
115
0.76
81.
239
0.07
3U
CM
0.01
80.
370
0.23
41.
695
0.05
40.
004
0.12
30.
746
1.34
60.
063
(e)α
1=
0.5,β
0=
0.2,φ
1=
0.25
,µ
2=
0.25
,σ
2 η1
=0.
5(f
)α
1=
0.5,β
0=
0.2,φ
1=
0.5,µ
2=
0,σ
2 η1
=0.
5
50U
CU
−0.
014
>10
0.27
62.
042
0.21
8−
0.00
60.
495
0.36
21.
923
0.18
8U
CM
0.02
10.
766
−0.
391
2.46
30.
214
0.06
20.
876
−0.
416
2.50
30.
162
100
UC
U0.
000
0.29
00.
514
1.62
70.
146
−0.
003
0.25
80.
578
1.58
20.
124
UC
M0.
009
0.50
20.
070
1.91
50.
139
0.03
21.
224
−0.
252
2.27
20.
099
250
UC
U0.
000
0.16
60.
684
1.34
70.
089
0.00
20.
147
0.72
11.
307
0.08
4U
CM
0.01
20.
281
0.42
81.
527
0.07
20.
044
0.57
5−
0.08
02.
168
0.06
0
(g)α
1=
0.5,β
0=
0.2,φ
1=
0.5,µ
2=
0.25
,σ
2 η1
=0.
25(h
)α
1=
0.5,β
0=
0,φ
1=
0.5,µ
2=
0.25
,σ
2 η1
=0.
75
50U
CU
−0.
009
0.24
00.
561
1.51
30.
233
−0.
002
1.00
60.
379
1.89
40.
196
UC
M0.
003
0.34
80.
350
1.68
90.
218
0.01
51.
935
−0.
176
2.22
60.
187
100
UC
U0.
000
0.14
70.
721
1.30
60.
142
0.00
60.
243
0.57
11.
506
0.13
0U
CM
0.00
30.
224
0.56
81.
435
0.16
40.
011
0.42
80.
257
1.75
90.
117
250
UC
U0.
000
0.08
40.
830
1.16
90.
078
0.00
20.
139
0.72
41.
288
0.08
2U
CM
0.00
50.
132
0.73
31.
244
0.07
10.
011
0.22
90.
529
1.42
70.
064
Note
s:See
Table
1.
16
Table
5:
Monte
Carl
ore
sult
sfo
rca
seII
I:λ
2=
1,σ
2 ε1
=1,σ
2 ε2
=1
σε1ε2
=0,σ
ε1η1
=0,σ
ε2η1
=0.5
TE
stim
ator
bias
(λ2)
Rm
se(λ
2)
p2,5
%p97,5
%P
(λ2)
bias
(λ2)
Rm
se(λ
2)
p2,5
%p97,5
%P
(λ2)
(a)α
1=
0.25
,β
0=
0.2,φ
1=
0.5,µ
2=
0.25
,σ
2 η1
=0.
5(b
)α
1=
0.5,β
0=
0.2,φ
1=
0.5,µ
2=
0.25
,σ
2 η1
=0.
5
50U
CU
0.16
00.
232
0.84
91.
496
0.32
10.
230
0.38
50.
770
1.82
60.
296
UC
M0.
062
0.29
40.
372
1.49
80.
220
0.10
43.
193
0.04
11.
901
0.20
5
100
UC
U0.
161
0.19
10.
959
1.35
70.
427
0.23
10.
281
0.91
41.
548
0.38
6U
CM
0.03
10.
213
0.53
91.
315
0.13
60.
053
0.34
30.
315
1.49
50.
133
250
UC
U0.
163
0.17
31.
044
1.28
20.
779
0.23
20.
250
1.04
51.
417
0.70
8U
CM
0.01
50.
120
0.74
01.
205
0.07
20.
026
0.18
10.
614
1.31
20.
070
(c)α
1=
0.75
,β
0=
0.2,φ
1=
0.5,µ
2=
0.25
,σ
2 η1
=0.
5(d
)α
1=
0.5,β
0=
0,φ
1=
0.5,µ
2=
0.25
,σ
2 η1
=0.
5
50U
CU
0.43
1>
100.
573
3.20
60.
247
0.22
40.
372
0.76
21.
807
0.29
6U
CM
0.21
32.
966
−0.
764
3.35
60.
173
0.10
60.
494
0.06
71.
883
0.20
6
100
UC
U0.
440
>10
0.81
92.
214
0.30
90.
230
0.28
10.
907
1.55
00.
370
UC
M0.
109
0.73
7−
0.32
32.
048
0.09
90.
054
0.34
60.
339
1.49
80.
130
250
UC
U0.
447
0.49
61.
082
1.85
10.
671
0.23
10.
249
1.04
11.
417
0.69
6U
CM
0.05
10.
363
0.23
01.
619
0.06
50.
025
0.18
00.
619
1.31
10.
071
(e)α
1=
0.5,β
0=
0.2,φ
1=
0.25
,µ
2=
0.25
,σ
2 η1
=0.
5(f
)α
1=
0.5,β
0=
0.2,φ
1=
0.5,µ
2=
0,σ
2 η1
=0.
5
50U
CU
0.35
90.
584
0.76
22.
193
0.31
50.
356
3.15
30.
796
2.21
20.
369
UC
M0.
166
0.69
8−
0.27
82.
371
0.21
50.
319
1.47
5−
0.05
22.
425
0.26
0
100
UC
U0.
366
0.20
50.
931
1.81
60.
455
0.36
00.
435
0.97
31.
851
0.51
8U
CM
0.08
70.
407
0.04
01.
756
0.13
90.
301
0.66
4−
0.14
02.
158
0.19
0
250
UC
U0.
366
0.39
01.
109
1.63
20.
805
0.35
50.
381
1.12
21.
609
0.86
6U
CM
0.04
00.
268
0.42
11.
465
0.06
80.
286
0.60
8−
0.06
41.
964
0.12
2
(g)α
1=
0.5,β
0=
0.2,φ
1=
0.5,µ
2=
0.25
,σ
2 η1
=0.
25(h
)α
1=
0.5,β
0=
0,φ
1=
0.5,µ
2=
0.25
,σ
2 η1
=0.
75
50U
CU
0.14
70.
255
0.78
81.
566
0.28
50.
296
1.15
70.
759
2.03
20.
311
UC
M0.
069
0.56
20.
326
1.61
00.
230
0.13
10.
652
−0.
143
2.14
00.
199
100
UC
U0.
152
0.19
40.
910
1.38
90.
325
0.29
70.
354
0.92
31.
673
0.42
9U
CM
0.03
60.
226
0.52
11.
351
0.15
80.
066
0.41
20.
179
1.61
80.
117
250
UC
U0.
156
0.17
01.
019
1.29
20.
611
0.29
10.
314
1.07
41.
519
0.76
9U
CM
0.01
60.
128
0.72
41.
219
0.08
10.
034
0.22
10.
526
1.38
20.
070
Note
s:See
Table
1.
17
4 An empirical application: testing PPP
The PPP proposition states that once converted to a common currency, national price levels should
be equal
pt − st = p∗t , (29)
where pt is the log of the price level in the home country, st is the log of the nominal exchange
rate, measured as the home currency price of one unit of foreign exchange, and p∗t is the log of the
price level in the foreign country. Defining the real exchange rate qt as the relative price of foreign
goods
qt = st − pt + p∗t , (30)
PPP holds if qt = 0. While numerous empirical studies have shown that PPP does not hold in the
short run (see e.g. Frenkel, 1981), the question whether PPP serves as an anchor for the long-run
real exchange rate has been an area of animated research. Empirical studies in the 1980s typically
concluded that PPP failed to hold even in the long run as (i) qt was found to follow a random
walk (see e.g. Adler and Lehman, 1983) and (ii) even after relaxing the assumption of long-run
homogeneity st, pt and p∗t were not found to be cointegrated (see e.g. Corbae and Ouliaris, 1988).
This failure to find evidence in favour of PPP is often attributed to low power of standard (ADF
type) unit root and cointegration tests to stationary alternatives with a root close to unity. Using
longer data series and higher-powered techniques, more recent tests do find evidence in favour of
long-run PPP (see e.g. Kim, 1990; Ardeni and Lubian, 1991; Glen, 1992; Taylor, 2002). Using
data on the real exchange rate, relative to the U.S. Dollar, for 19 countries over a period of more
than 100 years Taylor (2002), for instance, shows that the unit root hypothesis can be rejected for
most countries using a generalised-least-squares (GLS) version of the DF unit root test.
Engel (2000) argues that these unit root and cointegration tests in favour of PPP may have
reached the wrong conclusion, though. To see why, consider the price indices pt and p∗t to be
weighted averages of traded and non-traded goods prices
pt = (1− κ)pTt + κpN
t ,
p∗t = (1− κ∗)pT∗t + κ∗pN∗
t ,
where the superscripts T and N indicate traded and non-traded goods respectively and κ and
κ∗ are the shares that non-traded goods take in the overall price index in the home and the
foreign country respectively. The real exchange rate can now be decomposed into the relative
price of traded goods qTt and a component qN
t which is a weighted difference of the relative price
of non-traded to traded goods prices
qt = qTt + qN
t , (31)
18
where
qTt = st + pT∗
t − pTt , (32)
qNt = κ∗(pN∗
t − pT∗t )− κ(pN
t − pTt ). (33)
As almost any theory of international price determination implies that deviations from the law
of one price for traded goods are stationary, Engel (2000) argues that if qt is found to be non-
stationary this should be because the relative price of non-traded goods qNt is non-stationary,
which can be due to permanent shocks to productivity for instance. As there are no reliable long
time series available for qNt , non-stationarity of qN
t is tested using unit root tests on qt. This
hypothesis has been rejected in the recent literature. Using a simulation experiment, Engel (2000)
shows that both unit root and ECM cointegration tests on qt have a considerable size bias in
the presence of a (large) permanent component in qt, i.e. for a nominal size of 5% the true size
ranges from 90% to 99% in 100-year long data series when there is a non-stationary component
that accounts for 42% of the 100-year forecast variance of qt. This size bias is confirmed by the
simulation results in the previous section.
Combining equations (30) and (31)
pt − st = p∗t − qTt − qN
t , (34)
shows that there is a one-to-one relation between pt−st and p∗t but that this is not a cointegrating
relation if qNt is non-stationary. As standard estimators yield spurious results in this case, we will
estimate the long-run relation between pt− st and p∗t using the UC framework outlined in section
2.4. The model in equation (32)-(34) can be written is the format of the model in equations (1)-(4)
by setting y1t = pt − st, y2t = p∗t and assuming (see equation (6))
qTt = − (γ1(L)∆y1t + γ2(L)∆y2t + ε1t/α(1)) , (35)
∆qNt = −η1t/α(1). (36)
Consistent with the assumptions in Engel (2000), equation (35) allows qTt to be a stationary
process with rich dynamics while equation (36) restricts qNt to be a simple random walk. The
covariance matrix Ω in equation (5) allows shocks to the observed variables pt − st and p∗t and to
the unobserved component qNt to be contemporaneously correlated.
Data are taken from Taylor (2002). They consist of the log of annual nominal exchange rates
sit measured as domestic currency units per U.S. Dollar and the log of consumer price deflators pit,
where the index i = 1, . . . , 20 covers a set of 20 countries and the index t = 1892, . . . , 1996 covers
a set of 105 years. The variables y1t and y2t in equations (1)-(2) are taken to be the U.S. Dollar-
denominated price levels in the domestic country, pit−sit, and abroad, p∗it, respectively. In order to
avoid the choice of a base country, p∗it is constructed by aggregating for each country i the price lev-
els of the other 19 countries in the dataset into a ‘world’ basket as p∗it = (N − 1)−1 ∑j 6=i (pjt − sjt).
19
Given that p∗it is a ‘world’ basket, the model’s assumption that y2t is weakly exogenous for the
long-run parameters seems reasonable.
A standard ADF unit root test shows that the unit root hypothesis cannot be rejected for
pit − sit and p∗it in any of the included countries (results available on request). As a first step,
we therefore estimate the long-run relation between pit − sit and p∗it using the OLS, DGLS and
ECM estimators and test for cointegration using an ADF test on the residuals of the OLS and
DGLS regression and a t-test on the error-correction term in the ECM regression. Panel (a) in
Table 6 reports the results for the ECM estimator (the results for the OLS and DGLS estimators
are qualitatively similar and therefore not reported). At the 5% level of significance, the null
hypothesis of no cointegration is rejected for only 10 out of the 20 countries. Moreover, using a
standard t-test λ2 is found to be significantly different from 1 in 4 out of the 10 countries where
cointegration is found. So, the evidence in favour of PPP is at most modest.
Next, we estimate the long-run relation between pit − sit and p∗it using the UC model in
equations (15)-(16). As the full model is not necessarily identified, we first estimate a univariate
AR model to check whether p∗t has dynamics that are sufficiently rich for identification.3 As using
an ADF unit root test p∗t is found to be non-stationary, we impose a unit root by estimating the
model in first differences. The results (standard errors in brackets)
∆p∗t =0.026
(0.009) +0.463
(0.097) ∆p∗t−1 −0.258
(0.097) ∆p∗t−2. (37)
indicate that p∗t is a non-stationary AR(3) model with drift, which is sufficiently rich for identi-
fication. The results of applying the UC M estimator are reported in Panel (b) of Table 6. The
most important result is that the PPP hypothesis is rejected in only 3 out of 20 countries, i.e.
Canada, Denmark and Japan. While λ2 is not too far from 1 for Canada and Denmark, it equals
1.47 for Japan. The strong rejection of the PPP hypothesis in Japan may be due to the fact that
the secular increase in the relative price of non-traded goods induces a drift in the unobserved
component, which is not allowed for in the model. Note that out of the 10 countries for which we
could not reject the null hypothesis of no cointegration based on the ECM, 7 countries (Brazil,
Germany, the Netherlands, Portugal, Spain, Sweden and Switzerland) have ση1 6= 0. Further note
that in 2 countries (Finland and Mexico) for which we found cointegration based on the ECM,
we also have ση1 6= 0. This indicates that in these countries the conclusion that pit − sit and
p∗it are cointegrated is wrong. Consistent with the line of argumentation in Engel (2000), the
UC framework shows that U.S. Dollar-denominated price levels converge to their PPP level in
most countries once a permanent component in the real exchange rate, which may be induced by
non-stationarity of the relative price of non-traded goods, is modelled explicitly.
3Note that the ‘world’ basket p∗it is slightly different for each country i. Therefore, in this explorative estimation
p∗t is calculated as the average over the 20 countries in the sample, i.e. p∗t = (N)−1 Pj (pjt − sjt).
20
Table
6:
Tes
ting
the
PP
Phypoth
esis
usi
ng
the
EC
Mand
UC
Mes
tim
ato
rs
(a)
EC
M(b
)U
CM
pq
λ2
t CI
pq
α(1
)λ
2σ
2 ε1
σ2 ε2
σ2 η1
σε1ε2
σε2η1
Q(1
)Q
(4)
Arg
enti
na1
10.
92(0.0
8)−
5.18
∗∗∗
11
−0.
45(0.0
8)0.
91(0.0
9)0.
096
0.00
50.
000
−0.
003
0.00
00.
840.
32A
ustr
alia
11
0.92
(0.0
5)−
3.17
∗1
1−
0.17
(0.0
5)0.
92(0.0
5)0.
005
0.00
60.
000
−0.
001
0.00
00.
290.
67B
elgi
um1
21.
10(0.0
6)−
3.93
∗∗1
1−
0.34
(0.0
7)1.
08(0.0
8)0.
030
0.00
60.
000
0.00
50.
000
0.07
0.29
Bra
zil
11
1.08
(0.2
7)−
2.18
11
−0.
89(0.2
6)0.
96(0.8
0)0.
010
0.00
60.
006
−0.
005
−0.
003
0.90
0.02
Can
ada
21
0.89
(0.0
4)−
3.54
∗∗2
1−
0.15
(0.0
4)0.
88(0.0
5)0.
002
0.00
70.
000
0.00
00.
000
0.35
0.91
Den
mar
k2
21.
13(0.0
4)−
4.63
∗∗∗
22
−0.
28(0.0
5)1.
13(0.0
4)0.
008
0.00
60.
000
−0.
002
0.00
00.
350.
48Fin
land
21
0.95
(0.0
4)−
4.81
∗∗∗
21
−0.
70(0.2
1)0.
99(0.1
7)0.
018
0.00
60.
002
0.00
60.
002
0.64
0.92
Fran
ce1
10.
94(0.0
4)−
3.41
∗∗1
1−
0.20
(0.0
5)0.
94(0.0
5)0.
005
0.00
60.
000
0.00
10.
000
0.14
0.22
Ger
man
y2
11.
03(0.1
0)−
2.04
11
−0.
74(0.1
1)0.
92(0.2
1)0.
000
0.00
60.
004
0.00
1−
0.00
10.
620.
29It
aly
11
0.96
(0.0
9)−
2.90
11
−0.
19(0.0
6)0.
95(0.1
1)0.
023
0.00
70.
000
−0.
006
0.00
00.
230.
42Ja
pan
21
1.47
(0.0
5)−
3.62
∗∗2
1−
0.21
(0.0
5)1.
47(0.0
5)0.
006
0.00
60.
000
0.00
00.
000
0.67
0.86
Mex
ico
11
0.76
(0.0
6)−
4.46
∗∗∗
21
−0.
78(0.1
6)0.
76(0.2
4)0.
019
0.00
60.
004
0.00
20.
002
0.94
1.00
Net
herl
ands
12
1.13
(0.0
9)−
1.77
11
−0.
81(0.1
0)1.
02(0.1
5)0.
000
0.00
60.
004
0.00
1−
0.00
20.
780.
96N
orw
ay2
21.
04(0.0
4)−
3.42
∗∗2
1−
0.20
(0.0
6)1.
02(0.0
4)0.
007
0.00
60.
000
0.00
40.
000
0.34
0.88
Por
tuga
l1
20.
86(0.0
6)−
3.39
∗1
2−
0.51
(0.1
1)0.
88(0.1
4)0.
016
0.00
60.
001
0.00
70.
000
0.43
0.71
Spai
n1
10.
91(0.1
1)−
1.91
11
−1.
08(0.1
5)1.
11(0.2
0)0.
000
0.00
60.
006
0.00
00.
002
0.20
0.11
Swed
en1
10.
98(0.0
5)−
2.49
11
−1.
03(0.1
2)1.
00(0.1
4)0.
000
0.00
60.
003
−0.
001
0.00
00.
750.
99Sw
itze
rlan
d1
11.
27(0.0
6)−
2.41
11
−0.
83(0.1
0)1.
21(0.1
8)0.
000
0.00
60.
004
0.00
1−
0.00
20.
250.
75U
K1
10.
93(0.0
4)−
3.00
11
−0.
18(0.0
5)0.
93(0.0
5)0.
003
0.00
60.
000
0.00
10.
000
0.40
0.21
US
21
0.98
(0.0
3)−
4.45
∗∗∗
21
−0.
15(0.0
3)0.
96(0.0
5)0.
001
0.00
70.
000
0.00
00.
000
0.87
0.49
Note
:Sta
ndard
erro
rsin
bra
cket
s.*
den
ote
ssi
gnifi
cance
at
the
10%
level
.**
den
ote
ssi
gnifi
cance
at
the
5%
level
.***
den
ote
ssi
gnifi
cance
at
the
1%
level
.T
he
lag
ord
ers
pand
qin
the
EC
Mand
UC
Mare
chose
nby
sett
ing
am
axim
um
lag
ord
erof
2and
del
etin
gin
signifi
cant
lags
com
bin
edw
ith
ast
andard
test
for
seri
alco
rrel
ation
inth
ere
siduals
base
don
the
Box-L
jung
stati
stic
.T
he
lag
ord
err
inth
eU
CM
isse
tto
2(s
eere
sult
seq
uation
(37))
.For
the
EC
M,t C
Iis
ate
stfo
rth
enull
ofno
coin
tegra
tion,
i.e.
ast
andard
t-st
atist
icon
the
erro
r-co
rrec
tion
coeffi
cien
tbα(1)
inth
eE
CM
.T
he
sim
ula
ted
crit
icalvalu
esfo
rth
iste
stare
-4.0
3,-3
.41
and
-3.0
9at
the
1%
,5%
and
10%
level
resp
ecti
vel
y.For
the
UC
M,Q
(1)
and
Q(4
)are
p-v
alu
esofB
ox-L
jung
test
stati
stic
sfo
rse
rialco
rrel
ation
inth
eone-
step
fore
cast
erro
rsof
y2t
for
ala
gord
erof1
and
4re
spec
tivel
y.
21
5 Concluding comments
The concept of cointegration requires all variables constituting a long-run equilibrium relation to
be included in the analysis. Integrated variables that are among a set of variables constituting an
equilibrium relation but that are omitted from the analysis imply that (i) the remaining variables
are not cointegrated and (ii) standard cointegration tests have a serious size bias. As such, standard
estimators for a long-run relation between integrated variables may yield spurious results with
standard cointegration tests indicating these results to be a cointegration regression. This paper
uses an UC framework to estimate the long-run relationship between integrated variables when
possibly not all relevant variables are included in the analysis. A Monte Carlo experiment shows
that for a simple dynamic bivariate triangular process, this approach yields results similar to
those of standard estimators in the case of cointegration and does not yield spurious results in the
cases where the included variables are independent random walks or related in the long-run with an
unobserved component. Based on a Monte Carlo simulation, Inder (1993) concludes that ‘estimates
which include dynamics are much more reliable, even if the dynamic structure is overspecified’. The
Monte Carlo simulation in this paper suggests that estimates that include unobserved integrated
components may be even more reliable, even if both the dynamic structure and the error structure
are overspecified. As the general UC model allows shocks to the observed and to the unobserved
variables to be contemporaneously correlated, one important issue is that of identification. In
this paper, identification stems from the assumption that the unobserved component is a simple
random walk, with the observed explanatory variable having richer dynamics. In practise, this
assumption is not necessarily fulfilled. Allowing for richer dynamics in the unobserved component
is possible although for the model to be identified they should differ from the type of dynamics in
the observed components.
The proposed methodology is applied to testing PPP. Using data on consumer prices and
nominal exchange rates for a set of 20 countries over 105 years taken from Taylor (2002), standard
cointegration analysis provides only weak evidence in favour of PPP. Consistent with the line
of argumentation in Engel (2000), the UC framework shows that U.S. Dollar-denominated price
levels converge to their PPP level in most countries once a permanent component in the real
exchange rate, which may be induced by non-stationarity of the relative price of non-traded goods,
is modelled explicitly. This suggests that PPP holds for traded goods. The strong rejection of
the PPP hypothesis in Japan suggests that the assumption that the unobserved component is
a random walk may be too simplistic in some countries. A more detailed analysis with richer
dynamics in the unobserved component is left for future research.
22
References
Adler, M., Lehman, B., 1983. Deviations from purchasing power parity in the long run. Journal
of Finance 38 (5), 1471–1487.
Ardeni, P. G., Lubian, D., July 1991. Is there trend reversion in purchasing power parity? Euro-
pean Economic Review 35 (5), 1035–1055.
Banerjee, A., Dolado, J. J., Hendry, D. F., Smith, G. W., Aug. 1986. Exploring equilibrium rela-
tionships in econometrics through static models: Some monte-carlo evidence. Oxford Bulletin
of Economics and Statistics 48 (3), 253–277.
Banerjee, A., Donald, J., Mestre, R., 1996. Ecm tests for cointegration in a single equation frame-
work. Tech. rep., Oxford University.
Corbae, D., Ouliaris, S., August 1988. Cointegration and tests of purchasing power parity. The
Review of Economics and Statistics 70 (3), 508–11.
Durbin, J., Koopman, S., 2001. Time Series Analysis by State Space Methods. Oxford University
Press, Oxford.
Engel, C., August 2000. Long-run ppp may not hold after all. Journal of International Economics
51 (2), 243–273.
Engle, R., Hendry, D., Richard, J., 1983. Exogeneity. Econometrica 51, 277–304.
Engle, R. F., Granger, C. W. J., March 1987. Co-integration and error correction: Representation,
estimation, and testing. Econometrica 55 (2), 251–76.
Ericsson, N. R., MacKinnon, J. G., 2002. Distributions of error correction tests for cointegration.
Econometrics Journal 5 (2), 285–318.
Frenkel, J. A., 1981. The collapse of purchasing power parities during the 1970s. European Eco-
nomic Review 16 (1), 145–165.
Glen, J. D., Aug. 1992. Real exchange rates in the short, medium and long run. Journal of
International Economics 33 (1-2), 147–166.
Granger, C., Newbold, P., 1986. Forecasting Economic Time Series, 2nd Edition. Academic Press,
Orlando.
Harris, D., Inder, B., 1994. A test of the null of cointegration. In: Hargreaves, C. (Ed.), Non-
Stationary Time Series Analysis and Cointegration. Oxford University Press, Oxford, pp. 133–
152.
23
Harvey, A., 1989. Forecasting structural time series models and the Kalman filter. Cambridge
University Press.
Harvey, A. C., Henry, S. G. B., Peters, S., Wren-Lewis, S., Dec. 1986. Stochastic trends in dy-
namic regression models: An application to the employment-output equation. Economic Journal
96 (384), 975–985.
Inder, B., 1993. Estimating long-run relationships in economics: A comparison of different ap-
proaches. Journal of Econometrics 57 (1-3), 53–68.
Johansen, S., 1988. Statistical analysis of cointegration vectors. Journal of Economic Dynamics &
Control 12 (2-3), 231–254.
Kim, Y. B., Nov. 1990. Purchasing power parity in the long run: A cointegration approach. Journal
of Money, Credit and Banking 22 (4), 491–503.
Kremers, J. J. M., Ericsson, N. R., Dolado, J. J., August 1992. The power of cointegration tests.
Oxford Bulletin of Economics and Statistics 54 (3), 325–48.
Kwiatkowski, D., Phillips, P. C. B., Schmidt, P., Shin, Y., 1992. Testing the null hypothesis of
stationarity against the alternative of a unit root : How sure are we that economic time series
have a unit root? Journal of Econometrics 54 (1-3), 159–178.
Leybourne, S. J., McCabe, B. P. M., February 1994. A simple test for cointegration. Oxford
Bulletin of Economics and Statistics 56 (1), 97–103.
Mackinnon, J. G., 1996. Numerical distribution functions for unit root and cointegration tests.
Journal of Applied Econometrics 11 (6), 601–618.
Morley, J. C., 2007. The slow adjustment of aggregate consumption to permanent income. Journal
of Money, Credit, and Banking (in press).
Morley, J. C., Nelson, C. R., Zivot, E., 2003. Why are the beveridge-nelson and unobserved-
components decompositions of gdp so different? Review of Economics and Statistics 85 (2),
235–243.
Nelson, C. R., Plosser, C. I., 1982. Trends and random-walks in macroeconomic time-series: Some
evidence and implications. Journal of Monetary Economics 10 (2), 139–162.
Pagan, A., 1980. Some identification and estimation rresults for regression models with stochasti-
cally varying coefficients. Journal of Econometrics 13 (3), 341–363.
Phillips, P. C. B., Durlauf, S. N., Aug. 1986. Multiple time-series regression with integrated
porcesses. Review of Economic Studies 53 (4), 473–495.
24
Phillips, P. C. B., Hansen, B. E., January 1990. Statistical inference in instrumental variables
regression with i(1) processes. Review of Economic Studies 57 (1), 99–125.
Rudd, J., Whelan, K., 2006. Empirical proxies for the consumption-wealth ratio. Review of Eco-
nomic Dynamics 9, 34–51.
Saikkonen, P., Mar. 1991. Asymptotically efficient estimation of cointegration regressions. Econo-
metric Theory 7 (1), 1–21.
Sarantis, N., Stewart, C., 2001. Unobserved components in an error-correction model of consump-
tion for southern european countries. Empirical Economics 26, 391–405.
Shin, Y. C., Mar. 1994. A residual-based test of the null of cointegration against the alternative
of no cointegration. Econometric Theory 10 (1), 91–115.
Stock, J., September 1987. Asymptotic properties of least squares estimators of cointegrating
vectors. Econometrica 55 (5), 1035–56.
Stock, J. H., Watson, M. W., July 1993. A simple estimator of cointegrating vectors in higher
order integrated systems. Econometrica 61 (4), 783–820.
Taylor, A. M., 2002. A century of purchasing-power parity. The Review of Economics and Statistics
84 (1), 139–150.
Zivot, E., 1994. Single equation conditional error correction model based tests for cointegration.
Discussion Papers in Economics at the University of Washington 94-12, Department of Eco-
nomics at the University of Washington.
25