# Explaining Cointegration Analysis: Part I

David F. Hendry

Nuﬃeld College, Oxford
and Katarina Juselius
European University Institute, Florence
September 10, 1999
Abstract
‘Classical’ econometric theory assumes that observed data come from a stationary process,
where means and variances are constant over time. Graphs of economic time series, and the his-
torical record of economic forecasting, reveal the invalidity of such an assumption. Consequently,
we discuss the importance of stationarity for empirical modeling and inference; describe the ef-
fects of incorrectly assuming stationarity; explain the basic concepts of non-stationarity; note
some sources of non-stationarity; formulate a class of non-stationary processes (autoregressions
with unit roots) that seem empirically relevant for analyzing economic time series; and show
when an analysis can be transformed by means of diﬀerencing and cointegrating combinations
so stationarity becomes a reasonable assumption. We then describe how to test for unit roots
and cointegration. Monte Carlo simulations and empirical examples illustrate the analysis.
1 Introduction
Much of ‘classical’ econometric theory has been predicated on the assumption that the observed data
come from a stationary process, meaning a process whose means and variances are constant over
time. A glance at graphs of most economic time series, or at the historical track record of economic
forecasting, suﬃces to reveal the invalidity of that assumption: economies evolve, grow, and change
over time in both real and nominal terms, sometimes dramatically – and economic forecasts are
often badly wrong, although that should occur relatively infrequently in a stationary process.
Figure 1 shows some ‘representative’ time series to emphasize this point.
1
The ﬁrst panel (denoted
a) reports the time series of broad money in the UK over 1868–1993 on a log scale, together with
the corresponding price series (the UK data for 1868–1975 are from Friedman and Schwartz, 1982,
extended to 1993 by Attﬁeld, Demery and Duck, 1995). From elementary calculus, since ∂ log y/∂y =
1/y, the log scale shows proportional changes: hence, the apparently small movements between the
minor tic marks actually represent approximately 50% changes. Panel b shows real (constant price)
money on a log scale, together with real (constant price) output also on a log scale: the wider spacing
of the tic marks reveals a much smaller range of variation, but again, the notion of a constant mean
seems untenable. Panel c records long-run and short-run interest rates in natural scale, highlighting
changes over time in the variability of economic series, as well as in their means, and indeed, of

Financial support from the U.K. Economic and Social Research Council under grant R000234954, and from the
Danish Social Sciences Research Council is gratefully acknowledged. We are pleased to thank Campbell Watkins for
helpful comments on, and discussion of, earlier drafts.
1
Blocks of four graphs are lettered notionally as a, b; c, d in rows from the top left; six graphs are a, b, c; d, e, f;
and so on.
1875 1900 1925 1950 1975 2000
5
7.5
10
12.5
(a)
UK nominal money
UK price level
1875 1900 1925 1950 1975 2000
7
8
9
(b)
UK real money
UK real output
1875 1900 1925 1950 1975 2000
.05
.1
.15
(c)
UK short−term interest rate
UK long−term interest rate
1950 1960 1970 1980 1990
0
.1
.2
(f)
UK annual inflation
US annual inflation
1960 1970 1980 1990
1
1.5
(d)
US real output
US real money
1950 1960 1970 1980 1990
.05
.1
.15
(e)
US short−term interest rate
US long−term interest rate
Figure 1: Some ‘representative’ time series
the relationships between them: all quiet till 1929, the two variables diverge markedly till the early
1950s, then rise together with considerable variation. Nor is the non-stationarity problem speciﬁc
to the UK: panels d and e show the comparative graphs to b and c for the USA, using post-war
quarterly data (from Baba, Hendry and Starr, 1992). Again, there is considerable evidence of change,
although the last panel f comparing UK and US annual inﬂation rates suggests the UK may exhibit
greater instability. It is hard to imagine any ‘revamping’ of the statistical assumptions such that
these outcomes could be construed as drawings from stationary processes.
2
Intermittent episodes of forecast failure (a signiﬁcant deterioration in forecast performance rel-
ative to the anticipated outcome) conﬁrm that economic data are not stationary: even poor models
of stationary data would forecast on average as accurately as they ﬁtted, yet that manifestly does
not occur empirically. The practical problem facing econometricians is not a plethora of congruent
models from which to choose, but to ﬁnd any relationships that survive long enough to be useful.
It seems clear that stationarity assumptions must be jettisoned for most observable economic time
series.
Four issues immediately arise:
1. how important is the assumption of stationarity for modeling and inference?
2. what are the eﬀects of incorrectly assuming it?
3. what are the sources of non-stationarity?
2
It is sometimes argued that economic time series could be stationary around a deterministic trend, and we will
comment on that hypothesis later.
2
4. can empirical analyses be transformed so stationarity becomes a valid assumption?
Essentially, the answers are ‘very’; ‘potentially hazardous’; ‘many and varied’; and ‘sometimes,
depending on the source of non-stationarity’. Only intuitive explanations for these answers can be
oﬀered at this stage of our paper, but roughly:
1. when data means and variances are non-constant, observations come from diﬀerent distribu-
tions over time, posing diﬃcult problems for empirical modeling;
2. assuming constant means and variances when that is false can induce serious statistical mis-
takes, as we will show;
3. non-stationarity can be due to evolution of the economy, legislative changes, technological
change, and political turmoil inter alia;
4. some forms of non-stationarity can be eliminated by transformations, and much of our paper
concerns an important case where that is indeed feasible.
We expand on all four of these issues below, but to develop the analysis, we ﬁrst discuss some
necessary econometric concepts based on a simple model (in Section 2), and deﬁne the properties
of a stationary and a non-stationary process (Section 3). As embryology often yields insight into
evolution, we next review the history of regression analyses with trending data (Section 4) and
consider the possibilities of obtaining stationarity by transformation (called cointegration). Section
5 uses simulation experiments to look at the consequences of data being non-stationary in regression
models, building on the famous study in Yule (1926). We then brieﬂy review tests to determine the
presence of non-stationarity in the class noted in Section 5 (called univariate unit-root processes:
Section 6), as well as the validity of the transformation in Section 4 (cointegration tests: Section 7).
Finally, we empirically illustrate the concepts and ideas for a data set consisting of gasoline prices
in two major locations (Section 8). Section 9 concludes.
Non-stationarity seems a natural feature of economic life. Legislative change is one obvious source
of non-stationarity, often inducing structural breaks in time series, but it is far from the only one.
Economic growth, perhaps resulting from technological progress, ensures secular trends in many
time series, as Figure 1 illustrated. Such trends need to be incorporated into statistical analyses,
which could be done in many ways, including the venerable linear trend. Our focus here will be on a
type of stochastic non-stationarity induced by persistent cumulation of past eﬀects, called unit-root
processes (an explanation for this terminology is provided below).
3
Such processes can be interpreted
as allowing a diﬀerent ‘trend’ at every point in time, so are said to have stochastic trends. Figure
2 provides an artiﬁcial example: ‘joining’ the successive observations as shown produces a ‘jumpy’
trend as compared to the (best-ﬁtting) linear trend.
There are many plausible reasons why economic data may contain stochastic trends. For example,
technology involves the persistence of acquired knowledge, so that the present level of technology
is the cumulation of past discoveries and innovations. Economic variables depending closely on
technological progress are therefore likely to have a stochastic trend. The impact of structural
changes in the world oil market is another example of non-stationarity. Other variables related to
3
Stochastic means the presence of a random variable.
3
1 2 3 4 5 6 7 8 9 10
2
4
6
8
10
Figure 2: An artiﬁcial variable with a stochastic trend
the level of any variable with a stochastic trend will ‘inherit’ that non-stationarity, and transmit
it to other variables in turn: nominal wealth and exports spring to mind, and therefore income
and expenditure, and so employment, wages etc. Similar consequences follow for every source of
stochastic trends, so the linkages in economies suggest that the levels of many variables will be
non-stationary, sharing a set of common stochastic trends.
A non-stationary process is, by deﬁnition, one which violates the stationarity requirement, so
its means and variances are non-constant over time. For example, a variable exhibiting a shift in
its mean is a non-stationary process, as is a variable with a heteroscedastic variance over time. We
will focus here on the non-stationarity caused by stochastic trends, and discuss its implications for
empirical modeling.
To introduce the basic econometric concepts, we consider a simple regression model for a variable
y
t
containing a ﬁxed (or deterministic) linear trend with slope β generated from an initial value y
0
by:
y
t
= y
0
+βt +u
t
for t = 1, ..., T. (1)
To make the example more realistic, the error term u
t
is allowed to be a ﬁrst-order autoregressive
process:
u
t
= ρu
t−1

t
. (2)
That is, the current value of the variable u
t
is aﬀected by its value in the immediately preceding
period (u
t−1
) with a coeﬃcient ρ, and by a stochastic ‘shock’ ε
t
. We discuss the meaning of the
autoregressive parameter ρ below. The stochastic ‘shock’ ε
t
is distributed as IN[0, σ
2
ε
], denoting an
4
independent (I), normal (N) distribution with a mean of zero (E[ε
t
] = 0) and a variance V[ε
t
] = σ
2
ε
:
since these are constant parameters, an identical distribution holds at every point in time. A
process such as ¦ε
t
¦ is often called a normal ‘white-noise’ process. Of the three desirable properties
of independence, identical distributions, and normality, the ﬁrst two are clearly the most important.
The process u
t
in (2) is not independent, so we ﬁrst consider its properties. In the following, we will
use the notation u
t
for an autocorrelated process, and ε
t
for a white-noise process. The assumption of
only ﬁrst-order autocorrelation in u
t
, as shown in (2), is for notational simplicity, and all arguments
generalize to higher-order autoregressions. Throughout the paper, we will use lower-case letters to
indicate logarithmic scale, so x
t
= log(X
t
).
Dynamic processes are most easily studied using the lag operator L (see e.g., Hendry, 1995,
chapter 4) such that Lx
t
= x
t−1
. Then, (2) can be written as u
t
= ρLu
t

t
or:
u
t
=
ε
t
1 −ρL
. (3)
When [ρ[ < 1, the term 1/(1 −ρL) in (3) can be expanded as (1 +ρL +ρ
2
L
2
+ ). Hence:
u
t
= ε
t
+ρε
t−1

2
ε
t−2
+ . (4)
Expression (4) can also be derived after repeated substitution in (2). It appears that u
t
is the sum
of all previous disturbances (shocks) ε
t−i
, but that the eﬀects of previous disturbances decline with
time because [ρ[ < 1. However, now think of (4) as a process directly determining u
t
– ignoring our
derivation from (2) – and consider what happens when ρ = 1. In that case, u
t
= ε
t

t−1

t−2
+ ,
so each disturbance persists indeﬁnitely and has a permanent eﬀect on u
t
. Consequently, we say
that u
t
has the ‘stochastic trend’ Σ
t
i=1
ε
i
. The diﬀerence between a linear stochastic trend and
a deterministic trend is that the increments of a stochastic trend are random, whereas those of
a deterministic trend are constant over time as Figure 2 illustrated. From (4), we notice that
ρ = 1 is equivalent to the summation of the errors. In continuous time, summation corresponds to
integration, so such processes are also called integrated, here of ﬁrst order: we use the shorthand
notation u
t
∼ I(1) when ρ = 1, and u
t
∼ I(0) when [ρ[ < 1.
From (4), when [ρ[ < 1, we can derive the properties of u
t
as:
E[u
t
] = 0 and V[u
t
] =
σ
2
ε
1 −ρ
2
. (5)
Hence, the larger the value of ρ, the larger the variance of u
t
. When ρ = 1, the variance of u
t
becomes indeterminate and u
t
becomes a random walk. Interpreted as a polynomial in L, (3) has
a factor of 1 − ρL, which has a root of 1/ρ: when ρ = 1, (2) is called a unit-root process. While
there may appear to be many names for the same notion, extensions yield important distinctions:
for example, longer lags in (2) preclude u
t
being a random walk, and processes can be integrated of
order 2 (i.e., I(2)), so have several unit roots.
Returning to the trend regression example, by substituting (2) into (1) we get:
y
t
= βt +
ε
t
1 −ρL
+y
0
and by multiplying through the factor (1 −ρL):
(1 −ρL)y
t
= (1 −ρL)βt + (1 −ρL)y
0

t
. (6)
From (6), it is easy to see why the non-stationary process which results when ρ = 1, is often called
a unit-root process, and why an autoregressive error imposes a common-factor dynamics on a static
5
regression model (see e.g., Hendry and Mizon, 1978). When ρ = 1 in (6), the root of the lag
polynomial is unity, so it describes a linear diﬀerence equation with a unit coeﬃcient.
Rewriting (6) using Lx
t
= x
t−1
, we get:
y
t
= ρy
t−1
+β(1 −ρ)t +ρβ + (1 −ρ)y
0

t
, (7)
and it appears that the ‘static’ regression model (1) with autocorrelated residuals is equivalent to
the following dynamic model with white-noise residuals:
y
t
= b
1
y
t−1
+b
2
t +b
0

t
(8)
where
b
1
= ρ
b
2
= β(1 −ρ)
b
0
= ρβ + (1 −ρ)y
0
.
(9)
We will now consider four diﬀerent cases, two of which correspond to non-stationary (unit-root)
models, and the other two to stationary models:
Case 1. ρ = 1 and β ,= 0. It follows from (7) that ∆y
t
= β + ε
t
, for t = 1, . . . , T, where
∆y
t
= y
t
−y
t−1
. This model is popularly called a ‘random walk with drift’. Note that E[∆y
t
] = β ,= 0
is equivalent to y
t
having a linear trend, since although the coeﬃcient of t in (7) is zero, the coeﬃcient
of y
t−1
is unity so it ‘integrates’ the ‘intercept’ β, just as u
t
cumulated past ε
t
in (4).
Case 2. ρ = 1 and β = 0. From Case 1, it follows immediately that ∆y
t
= ε
t
. This is called a pure
random walk model: since E[∆y
t
] = 0, y
t
contains no linear trend.
Case 3. [ρ[ < 1 and β ,= 0 gives us (8), i.e., a ‘trend-stationary’ model. The interpretation of the
coeﬃcients b
1
, b
2
, and b
0
must be done with care: for example, b
2
is not an estimate of the trend in
y
t
, instead β = b
2
/(1 −ρ) is the trend in the process.
4
Case 4. [ρ[ < 1 and β = 0 delivers y
t
= ρy
t−1
+ (1 − ρ)y
0
+ ε
t
, which is the usual stationary
autoregressive model with a constant term.
Hence:
• in the static regression model (1), the constant term is essentially accounting for the unit of
measurement of y
t
, i.e., the ‘kick-oﬀ’ value of the y series;
• in the dynamic regression model (8), the constant term is a weighted average of the growth
rate β and the initial value y
0
;
• in the diﬀerenced model (ρ = 1), the constant term is solely measuring the growth rate, β.
We will now provide some examples of economic variables that can be appropriately described by
the above simple models.
3 Properties of a non-stationary and a stationary process
Unit-root processes can also arise as a consequence of plausible economic behavior. As an example,
we will discuss the possibility of a unit root in the long-term interest rate. Similar arguments
could apply to exchange rates, other asset prices, and prices of widely-traded commodities, such as
gasoline.
4
We doubt that GNP (say) has a deterministic trend, because of the thought experiment that output would then
continue to grow if we all ceased working.....
6
If changes to long-term interest rates (R
l
) were predictable, and R
l
> R
s
(the short-term rate)
– as usually holds, to compensate lenders for tying up their money – one could create a money
machine. Just predict the forthcoming change in R
l
, and borrow at R
s
to buy bonds if you expect a
fall in R
l
(a rise in bond prices) or sell short if R
l
is likely to rise. Such a scenario of boundless proﬁt
at low risk seems unlikely, so we anticipate that the expected value of the change in the long-term
interest rate at time t −1, given the relevant information set 1
t−1
, is zero, i.e., E
t−1
[∆R
l,t
[1
t−1
] = 0
(more generally, the change should be small on a risk-adjusted basis after transactions costs). As a
model, this translates into:
R
l,t
= R
l,t−1
+
t
(10)
where E
t−1
[
t
[1
t−1
] = 0 and
t
is an ID[0, σ
2
ε
] process (where D denotes the relevant distribution,
which need not be normal). The model in (10) has a unit coeﬃcient on R
l,t−1
, and as a dynamic
relation, is a unit-root process. To discuss the implications for empirical modeling of having unit
roots in the data, we ﬁrst need to discuss the statistical properties of stationary and non-stationary
processes.
5
1950 1960 1970 1980 1990
5
10
15
Long−term interest rate
1950 1960 1970 1980 1990
−1
0
1
Change in the long−term interest rate
0 5 10 15
.05
.1
.15
.2
Density
−1.5 −1 −.5 0 .5 1 1.5
1
2
3
Density
0 5 10
.5
1
Correlogram
0 5 10
0
1
Correlogram
Figure 3: Monthly long-term interest rates in the USA, in levels and diﬀerences, 1950–1993
3.1 A non-stationary process
Equation (10) shows that the whole of R
l,t−1
and
t
inﬂuence R
l,t
, and hence, in the next period,
the whole of R
l,t
inﬂuences R
l,t+1
and so on. Thus, the eﬀect of
t
persists indeﬁnitely, and past
5
Empirically, for the monthly data over 1950(1)–1993(12) on 20-year bond rates in the USA shown in Figure 3,
the estimated coeﬃcient of R
l,t−1
in (10) is 0.994 with an estimated standard error of 0.004.
7
errors accumulate with no ‘depreciation’, so an equivalent formulation of (10) is:
R
l,t
=
t
+
t−1
+ +
1
+
0
+
−1
(11)
or alternatively:
R
l,t
=
t
+
t−1
+ +
1
+R
l,0
(12)
where the initial value R
l,0
=
0
+
−1
... contains all information of the past behavior of the long-term
interest rate up to time 0. In practical applications, time 0 corresponds to the ﬁrst observation in the
sample. Equation (11) shows that theoretically, the unit-root assumption implies an ever-increasing
variance to the time series (around a ﬁxed mean), violating the constant-variance assumption of a
stationary process. In empirical studies, the conditional model (12) is more relevant as a description
of the sample variation, and shows that ¦R
l,t
[R
l,0
¦ has a ﬁnite variance, tσ
2

, but this variance is
non-constant since it changes with t = 1, . . . , T.
Cumulating random errors will make them smooth, and in fact, induces properties like those
of economic variables, as ﬁrst discussed by Working (1934) (so R
l,t
should be smooth, at least in
comparison to its ﬁrst diﬀerence, and is, as illustrated in Figure 3, panels a and b). From (12),
taking R
l,0
as a ﬁxed number, one can see that:
E[R
l,t
] = R
l,0
(13)
and that:
V[R
l,t
] = σ
2

t. (14)
Further, perhaps not so easily seen, when t > s, the covariance between drawings t −s periods apart
is:
C[R
l,t
, R
l,t−s
] = E[(R
l,t
−R
l,0
) (R
l,s
−R
l,0
)] = σ
2

s (15)
and so:
corr
2
[R
l,t
, R
l,t−s
] = 1 −
s
t
. (16)
Consequently, when t is large, all the serial correlations for a random-walk process are close to
unity, a feature of R
l,t
as illustrated in Figure 3, panel e. Finally, even if ¦R
l,t
¦ is the sum of
a large number of errors it will not be approximately normally distributed. This is because each
observation, R
l,t
[R
l,0
, t = 1, ..., T, has a diﬀerent variance. Figure 3, panel c shows the histogram
of approximately 80 quarterly observations for which the observed distribution is bimodal, and so
does not look even approximately normal. We will comment on the properties of ∆R
l,t
below.
To summarize, the variance of a unit-root process increases over time, and successive observations
are highly interdependent. The theoretical mean of the conditional process R
l,t
[R
l,0
is constant and
equal to R
l,0
. However, the theoretical mean and the empirical sample mean R
l
are not even
approximately equal when data are non-stationary (surprisingly, the sample mean divided by

3T
is distributed as N[0, 1] in large samples: see Hendry, 1995, chapter 3).
3.2 A stationary process
We now turn to the properties of a stationary process. As argued above, most economic time series
are non-stationary, and at best become stationary only after diﬀerencing. Therefore, we will from
the outset discuss stationarity either for a diﬀerenced variable ¦∆y
t
¦ or for the IID errors ¦ε
t
¦.
8
A variable ∆y
t
is weakly stationary when its ﬁrst two moments are constant over time, or more
precisely, when E[∆y
t
] = µ, E[(∆y
t
− µ)
2
] = σ
2
, and E[(∆y
t
− µ)(∆y
t−s
− µ)] = γ(s) ∀s, where µ,
σ
2
, and γ(s) are ﬁnite and independent of t.
6
As an example of a stationary process take the change in the long-term interest rate from (10):
∆R
l,t
=
t
where
t
∼ ID
_
0, σ
2

¸
. (17)
It is a stationary process with mean E[∆R
l,t
] = 0, E[(∆R
l,t
−0)
2
] = σ
2
ε
, and E[(∆R
l,t
−0)(∆R
l,t−s

0)] = 0 ∀s, because ∆R
l,t
= ε
t
is assumed to be an IID process. This is illustrated by the graph
of the change in the long-term bond rate in Figure 3, panel b. We note that the autocorrelations
have more or less disappeared (panel f), and that the density distribution is approximately normal
except for an outlier corresponding to the deregulation of capital movements in 1983 (panel d).
An IID process is the simplest example of a stationary process. However, as demonstrated in
(2), a stationary process can be autocorrelated, but such that the inﬂuence of past shocks dies out.
Otherwise, as demonstrated by (4), the variance would not be constant.
4 Regression with trending variables: a historical review
One might easily get the impression that the unit-root literature is a recent phenomenon. This is
clearly not the case: already in 1926, Udny Yule analyzed the hazards of regressing a trending variable
on another unrelated trending variable – the so-called ‘nonsense regression’ problem. However, the
development of ‘unit-root econometrics’ that aims to address this problem in a more constructive
way, is quite recent.
A detailed history of econometrics has also developed over the past decade and full coverage is
provided in Morgan (1990), Qin (1993), and Hendry and Morgan (1995). Here, we brieﬂy review
the evolution of the concepts and tools underpinning the analysis of non-stationarity in economics,
commencing with the problem of ‘nonsense correlations’, which are extremely high correlations often
found between variables for which there is no ready causal explanation (such as birth rates of humans
and the number of storks’ nests in Stockholm).
Yule (1926) was the ﬁrst to formally analyze such ‘nonsense correlations’. He thought that
they were not the result of both variables being related to some third variable (e.g., population
growth). Rather, he categorized time series according to their serial-correlation properties, namely,
how highly correlated successive values were with each other, and investigated how their cross-
correlation coeﬃcient r
xy
behaved when two unconnected series x and y had:
A] random levels;
B] random ﬁrst diﬀerences;
C] random second diﬀerences.
For example, in case B, the data take the form ∆x
t
= ε
t
(Case 2. in Section 3) where ε
t
is IID.
Since the value of x
t
depends on all past errors with equal weights, the eﬀects of distant shocks
persist, so the variance of x
t
increases over time, making it non-stationary. Therefore, the level of
x
t
contains information about all permanent disturbances that have aﬀected the variable, starting
from the initial level x
0
at time 0, and contains the linear stochastic trend Σε
i
. Similarly for the
y
t
series, although its stochastic trend depends on cumulating errors that are independent of those
entering x
t
.
6
When the moments depend on the initial conditions of the process, stationarity holds only asymptotically (see
e.g. Spanos, 1986), but we ignore that complication here.
9
Yule found that r
xy
was almost normally distributed in case A, but became nearly uniformly
distributed (except at the end points) in B. He was startled to discover that r
xy
distribution in C, so the correct null hypothesis (of no relation between x and y) was virtually
certain to be rejected in favor of a near-perfect positive or negative link. Consequently, it seemed
as if inference could go badly wrong once the data were non-stationary. Today, his three types of
series are called integrated of orders zero, one, and two respectively (I(0), I(1), and I(2) as above).
Diﬀerencing an I(1) series delivers an I(0), and so on. Section 5 replicates a simulation experiment
that Yule undertook.
Yule’s message acted as a signiﬁcant discouragement to time-series work in economics, but gradu-
ally its impact faded. However, Granger and Newbold (1974) highlighted that a good ﬁt with
signiﬁcant serial correlation in the residuals was a symptom associated with nonsense regressions.
Hendry (1980) constructed a nonsense regression by using cumulative rainfall to provide a better
explanation of price inﬂation than did the money stock in the UK. A technical analysis of the sources
and symptoms of the nonsense-regressions problem was ﬁnally presented by Phillips (1986).
As economic variables have trended over time since the Industrial Revolution, ensuring non-
stationarity resulted in empirical economists usually making careful adjustments for factors such as
population growth and changes in the price level. Moreover, they often worked with the logarithms
of data (to ensure positive outcomes and models with constant elasticities), and thereby implicitly
assumed constant proportional relations between non-stationary variables. For example, if β ,= 0
and u
t
is stationary in the regression equation:
y
t
= β
0

1
x
t
+u
t
(18)
then y
t
and x
t
must contain the same stochastic trend, since otherwise u
t
could not be stationary.
Assume that y
t
is aggregate consumption, x
t
is aggregate income, and the latter is a random walk,
i.e., x
t
= Σε
i
+ x
0
.
7
If aggregate income is linearly related to aggregate consumption in a causal
way, then y
t
would ‘inherit’ the non-stationarity from x
t
, and u
t
would be stationary unless there
were other non-stationary variables than income causing consumption.
Assume now that, as before, y
t
is non-stationary, but it is not caused by x
t
determined by another non-stationary variable, say, z
t
= Σν
i
+ z
0
, unrelated to x
t
. In this case,
β
1
= 0 in (18) is the correct hypothesis, and hence what one would like to accept in statistical tests.
Yule’s problem in case B can be seen clearly: if β
1
were zero in (18), then y
t
= β
0
+ u
t
; i.e., u
t
contains Σν
i
so is non-stationary and, therefore, inconsistent with the stationarity assumption of
the regression model. Thus, one cannot conduct standard tests of the hypothesis that β
1
= 0 in
such a setting. Indeed, u
t
being autocorrelated in (18), with u
t
being non-stationary as the extreme
case, is what induces the non-standard distributions of r
xy
.
Nevertheless, Sargan (1964) linked static-equilibrium economic theory to dynamic empirical mod-
els by embedding (18) in an autoregressive-distributed lag model:
y
t
= b
0
+b
1
y
t−1
+b
2
x
t
+b
3
x
t−1

t
. (19)
The dynamic model (19) can also be formulated in the so-called equilibrium-correction form by
subtracting y
t−1
from both sides and subtracting and adding b
2
x
t−1
to the right-hand side of (19):
∆y
t
= α
0

1
∆x
t
−α
2
(y
t−1
−β
1
x
t−1
−β
0
) +ε
t
(20)
where α
1
= b
2
, α
2
= (1−b
1
), β
1
= (b
2
+b
3
)/(1−b
1
), and α
0

2
β
0
= b
0
. Thus, all coeﬃcients in (20)
can be derived from (19). Models such as (20) explain growth rates in y
t
by the growth in x
t
and the
7
Recall lower case letters are in logaritmic form.
10
past disequilibrium between the levels. Think of a situation where consumption changes (∆y
t
) as a
result of a change in income (∆x
t
), but also as a result of previous period’s consumption not being
in equilibrium (i.e., y
t−1
,= β
0

1
x
t−1
). For example, if previous consumption was too high, it has
to be corrected downwards, or if it was too low, it has to be corrected upwards. The magnitude of
the past disequilibrium is measured by (y
t−1
− β
1
x
t−1
− β
0
) and the speed of adjustment towards
this steady-state by α
2
.
Notice that when ε
t
, ∆y
t
and ∆x
t
are I(0), there are two possibilities:
• α
2
,= 0 and (y
t−1
−β
1
x
t−1
−β
0
) ∼ I(0), or
• α
2
= 0 and (y
t−1
−β
1
x
t−1
−β
0
) ∼ I(1).
The latter case can be seen from (19). When α
2
= 0, then ∆y
t
= α
0

1
∆x
t

t
, and by integrating
we get y
t
= α
1
x
t
+

ε
t
. Notice also that by subtracting β
1
∆x
t
from both sides of (20) and collecting
terms, we can derive the properties of the equilibrium error u
t
= (y −β
1
x −β
0
)
t
:
(y −β
1
x −β
0
)
t
= α
0
+ (α
1
−β
1
)∆x
t
+ (1 −α
2
)(y −β
1
x −β
0
)
t−1

t
(21)
or:
u
t
= ρu
t−1

0
+ (α
1
−β
1
)∆x
t

t
.
where ρ = (1 −α
2
).
8
Thus, the equilibrium error is an autocorrelated process; the higher the value
of ρ (equivalently the smaller the value of α
2
), the slower is the adjustment back to equilibrium,
and the longer it takes for an equilibrium error to disappear. If α
2
= 0, there is no adjustment, and
y
t
does not return to any equilibrium value, but drifts as a non-stationary variable. To summarize:
when α
2
,= 0 (so ρ ,= 1), the ‘equilibrium error’ u
t
= (y − β
1
x − β
0
)
t
is a stationary autoregressive
process.
I(1) ‘nonsense-regressions’ problems will disappear in (20) because ∆y
t
and ∆x
t
are I(0) and,
therefore, no longer trending. Standard t-statistics will be ‘sensibly’ distributed (assuming that ε
t
is
IID), irrespective of whether the past equilibrium error, u
t−1
, is stationary or not.
9
This is because a
stationary variable, ∆y
t
, cannot be explained by a non-stationary variable, and ´ α
2
· 0 if u
t−1
∼ I(1).
Conversely, when u
t−1
∼ I(0), then ´ α
2
measures the speed of adjustment with which ∆y
t
(corrects) towards each new equilibrium position.
Based on equations like (20), Hendry and Anderson (1977) noted that ‘there are ways to achieve
stationarity other than blanket diﬀerencing’, and argued that terms like u
t−1
would often be station-
ary even when the individual series were not. More formally, Davidson, Hendry, Srba and Yeo (1978)
introduced a class of models based on (20) which they called ‘error-correction’ mechanisms (denoted
ECMs). To understand the status of equations like (20), Granger (1981) introduced the concept of
cointegration where a genuine relation exists, despite the non-stationary nature of the original data,
thereby introducing the obverse of nonsense regressions. Further evidence that many economic time
series were better construed as non-stationary than stationary was presented by Nelson and Plosser
(1982), who tested for the presence of unit roots and could not reject that hypothesis. Closing this
circle, Engle and Granger (1987) proved that ECMs and cointegration were actually two names for
8
The change in the equilibrium yt = β
1
xt − β
0
is ∆yt = β
1
∆xt, so these variables must have steady-state growth
rates gy and gx related by gy = β
1
gx. But from (20), gy = α
0
+ α
1
gx, hence we can derive that α
0
= −(α
1
− β
1
)gx,
as occurs in (21).
9
The resulting distribution is not actually a t-statistic as proposed by Student (1908): Section 6 shows that it
depends in part on the Dickey–Fuller distribution. However, t is well behaved, unlike the ‘nonsense regressions’ case.
11
the same thing: cointegration entails a feedback involving the lagged levels of the variables, and a
lagged feedback entails cointegration.
10
5 Nonsense regression illustration
Using modern software, it is easy to demonstrate nonsense regressions between unrelated unit-root
processes. Yule’s three cases correspond to generating uncorrelated bivariate time series, where the
data are cumulated zero, once and twice, before regressing either series on the other, and conducting
a t-test for no relationship. The process used to generate the I(1) data mimics that used by Yule,
namely:
∆y
t
=
t
where
t
∼ IN
_
0, σ
2

¸
(22)
∆x
t
= ν
t
where ν
t
∼ IN
_
0, σ
2
ν
¸
(23)
and setting y
0
= 0, x
0
= 0. Also:
E[
t
ν
s
] = 0 ∀t, s. (24)
The economic equation of interest is postulated to be:
y
t
= β
0

1
x
t
+u
t
(25)
where β
1
is believed to be the derivative of y
t
with respect to x
t
:
∂y
t
∂x
t
= β
1
.
Equations like (25) estimated by OLS wrongly assume ¦u
t
¦ to be an IID process independent of x
t
.
A t-test of H
0
: β
1
= 0 (as calculated by a standard regression package, say) is obtained by dividing
the estimated coeﬃcient by its standard error:
t
β
1
=0
=
´
β
1
SE
_
´
β
1
_ (26)
where:
´
β
1
=
_

(x
t
− x)
2
_
−1
(x
t
−x)(y
t
−y), (27)
and
SE
_
´
β
1
_
=
´ σ
u
_

(x
t
−x)
2
. (28)
When u
t
correctly describes an IID process, the t-statistic satisﬁes:
P

¸
t
β
1
=0
¸
¸
≥ 2.0 [ H
0
_
· 0.05. (29)
10
Some references to the vast literature on the theory and practice of testing for both unit roots and cointegration
include: Banerjee and Hendry (1992), Banerjee, Dolado, Galbraith and Hendry (1993), Chan and Wei (1988), Dickey
and Fuller (1979, 1981), Hall and Heyde (1980), Hendry (1995), Johansen (1988, 1991, 1992a, 1992b, 1995b), Johansen
and Juselius (1990, 1992), Phillips (1986, 1987a, 1987b, 1988), Park and Phillips (1988, 1989), Phillips and Perron
(1988), and Stock (1987).
12
This, however, is not the case if u
t
is autocorrelated: in particular, if u
t
is I(1). In fact, at T = 100,
from the Monte Carlo experiment, we would need a critical value of 14.8 to deﬁne a 5% rejection
frequency under the null because:
P

¸
t
β
1
=0
¸
¸
≥ 14.8 [ H
0
_
· 0.05,
so serious over-rejection occurs using (29). Instead of the conventional critical value of 2, we should
use 15.
Why does such a large distortion occur? We will take a closer look at each of the components
in (26) and expand their formulae to see what happens when u
t
is non-stationary. The intuition is
that although
´
β
1
is an unbiased estimator of β
1
, so E[
´
β
1
] = 0, it has a very large variance, but the
calculated SE[
´
β
1
] dramatically under-estimates the true value.
From (28), SE[
´
β
1
] consists of two components, the residual standard error, ´ σ
u
, and the sum of
squares,

(x
t
−x)
2
. When β
1
= 0, the estimated residual variance ´ σ
2
u
will in general be lower than
σ
2
u
=

(y
t
−y)
2
/T. This is because the estimated value
´
β
1
is usually diﬀerent from zero (sometimes
widely so) and, hence, will produce smaller residuals. Thus:

(y
t
− ´ y
t
)
2

(y
t
−y)
2
, (30)
where ´ y
t
=
´
β
0
+
´
β
1
x
t
. More importantly, the sum of squares

(x
t
− x)
2
is not an appropriate
measure of the variance in x
t
when x
t
is non-stationary. This is so because x (instead of x
t−1
) is
a very poor ‘reference line’ when x
t
is trending, as is evident from the graphs in Figures 1 and 3,
and our artiﬁcial example 2. When the data are stationary, the deviation from the mean is a good
measure of how much x
t
has changed, whereas when x
t
is non-stationary, it is the deviation from
the previous value that measures the stochastic change in x
t
. Therefore:

(x
t
−x)
2
¸

(x
t
−x
t−1
)
2
. (31)
so both (30) and (31) work in the same direction of producing a serious downward bias in the
estimated value of SE[
´
β
1
].
It is now easy to understand the outcome of the simulation study: because the correct standard
error is extremely large, i.e., σ
β
1
¸ SE[
´
β
1
], the dispersion of
´
β
1
around zero is also large, big positive
and negative values both occur, inducing many big ‘t-values’.
Figure 4 reports the frequency distributions of the t-tests from a simulation study by PcNaive
(see Doornik and Hendry, 1998), using M = 10, 000 drawings for T = 50. The shaded boxes are for
±2, which is the approximate 95% conventional conﬁdence interval. The ﬁrst panel (a) shows the
distribution of the t-test on the coeﬃcient of x
t
in a regression of y
t
on x
t
when both variables are
white noise and unrelated. This is numerically very close to the correct distribution of a t-variable.
The second panel (denoted b, in the top row) shows the equivalent distribution for the nonsense
regression based on (22)–(26). The third panel (c, left in the lower row) is for the distribution of the
t-test on the coeﬃcient of x
t
in a regression of y
t
on x
t
, y
t−1
and x
t−1
when the data are generated
as unrelated stationary ﬁrst-order autoregressive processes. The ﬁnal panel (d) shows the t-test on
the equilibrium-correction coeﬃcient α
2
in (20) for data generated by a cointegrated process (so
α
2
,= 0 and β is known).
The ﬁrst and third panels are close to the actual distribution of Student’s t; the former is as
expected from statistical theory, whereas the latter shows that outcome is approximately correct in
dynamic models once the dynamics have been included in the equation speciﬁcation. The second
panel shows an outcome that is wildly diﬀerent from t, with a distributional spread so wide that
13
−20 −10 0 10 20 30
.025
.05
.075
nonsense
regression
−4 −2 0 2 4
.1
.2
.3
.4
2 unrelated
variables
white−noise
−4 −2 0 2 4
.1
.2
.3
.4
2 unrelated
stationary
variables
0 2 4 6 8
.1
.2
.3
.4
.5
cointegrated
relation
Figure 4: Frequency distributions of nonsense-regression t-tests
most of the probability lies outside the usual region of ±2. While the last distribution is not centered
on zero – because the true relation is indeed non-null – it is included to show that the range of the
distribution is roughly correct.
Thus, panels (c) and (d) showthat, by themselves, neither dynamics nor unit-root non-stationarity
induce serious distortions: the nonsense-regressions problem is due to incorrect model speciﬁcation.
Indeed, when y
t−1
and x
t−1
are added as regressors in case (b), the correct distribution results for
t
β
1
=0
, delivering a panel very similar to (c), so the excess rejection is due to the wrong standard error
being used in the denominator (which as shown above, is badly downwards biased by the untreated
residual autocorrelation).
Almost all available software packages contain a regression routine that calculates coeﬃcient
estimates, t-values, and R
2
based on OLS. Since the computer will calculate the coeﬃcients inde-
pendently of whether the variables are stationary or not (and without issuing a warning when they
are not), it is important to be aware of the following implications for regressions with trending
variables:
(i) Although E[
´
β
1
] = 0, nevertheless t
β
1
=0
diverges to inﬁnity as T increases, so that conventionally-
calculated critical values are incorrect (see Hendry, 1995, chapter 3).
(ii) R
2
cannot be interpreted as a measure of goodness-of-ﬁt.
The ﬁrst point means that one will too frequently reject the null hypothesis (β
1
= 0) when it is
true. Even in the best case, when β
1
,= 0, i.e., when y
t
and x
t
are causally related, standard t-tests
will be biased with too frequent rejections of a null hypothesis such as β
1
= 1, when it is true.
14
Hence statistical inference from regression models with trending variables is unreliable based on
standard OLS output. The second point will be further discussed and illustrated in connection with
the empirical analysis.
All this points to the crucial importance of always checking the residuals of the empirical model
for (unmodeled) residual autocorrelation. If autocorrelation is found, then the model should be
re-speciﬁed to account for this feature, because many of the conventional statistical distributions,
such as Student’s t, the F, and the χ
2
distributions become approximately valid once the model
is re-speciﬁed to have a white-noise error. So even though unit roots impart an important non-
stationarity to the data, reformulating the model to have white-noise errors is a good step towards
solving the problem; and transforming the variables to be I(0) will complete the solution.
6 Testing for unit roots
We have demonstrated that stochastic trends in the data are important for statistical inference. We
will now discuss how to test for the presence of unit roots in the data. However, the distinction
between a unit-root process and a near unit-root process need not be crucial for practical modeling.
Even though a variable is stationary, but with a root close to unity (say, ρ > 0.95), it is often a
good idea to act as if there are unit roots to obtain robust statistical inference. An example of this
is given by the empirical illustration in Section 8.
We will now consider unit-root testing in a univariate setting. Consider estimating β in the
autoregressive model:
y
t
= βy
t−1
+
t
where
t
∼ IN
_
0, σ
2

¸
(32)
under the null of β = 1 and y
0
= 0 (i.e., no determinist trend in the levels), using a sample of
size T. Because V[y
t
] = σ
2

t, the data second moments (like

T
t=1
y
2
t−1
) grow at order T
2
, so the
distribution of
´
β −β ‘collapses’ very quickly. Again, we can illustrate this by simulation, estimating
(32) at T = 25, 100, 400, and 1000. The four panels for the estimated distribution in Figure 5 have
been standardized to the same x-axis for visual comparison – and the convergence is dramatic. For
comparison, the corresponding graphs for β = 0.5 are shown in Figure 6, where second moments grow
at order T. Thus, to obtain a limiting distribution for
´
β −β which neither diverges nor degenerates
to a constant, scaling by T is required (rather than

T for I(0) data). Moreover, even after such a
scaling, the form of the limiting distribution is diﬀerent from that holding under stationarity.
The ‘t-statistic’ for testing H
0
: β = 1, often called the Dickey–Fuller test after Dickey and Fuller
(1979), is easily computed, but does not have a standard t-distribution. Consequently, conventional
critical values are incorrect, and using them can lead to over-rejection of the null of a unit root
when it is true. Rather, the Dickey–Fuller (DF) test has a skewed distribution with a long left tail,
making it hard to discriminate the null of a unit root from alternatives close to unity. The more
general test, the Augmented Dickey-Fuller test is deﬁned in the next section.
Unfortunately, the form of the limiting distribution of the DF test is also altered by the presence
of a constant or a trend in either the DGP or the model. This means that diﬀerent critical values
are required in each case, although all the required tables of the correct critical values are available.
Worse still, wrong choices of what deterministic terms to include – or which table is applicable –
can seriously distort inference. As demonstrated in Section 2, the role and the interpretation of
the constant term and the trend in the model changes as we move from the stationary case to the
non-stationary unit-root case. It is also the case for stationary data that incorrectly omitting (say)
an intercept can be disastrous, but mistakes are more easily made when the data are non-stationary.
15
−.25 0 .25 .5 .75 1 1.25
1
2
3
T=25
−.25 0 .25 .5 .75 1 1.25
5
10
T=100
−.25 0 .25 .5 .75 1 1.25
10
20
30
40
50
T=400
−.25 0 .25 .5 .75 1 1.25
50
100
T=1000
Figure 5: Frequency distributions of unit-root estimators
However, always including a constant and a trend in the estimated model ensures that the test will
have the correct rejection frequency under the null for most economic time series. The required
critical values have been tabulated using Monte Carlo simulations by Dickey and Fuller (1979,
1981), and most time-series econometric software (e.g., PcGive) automatically provides appropriate
critical values for unit-root tests in almost all relevant cases, provided the correct model is used (see
discussion in the next section).
7 Testing for cointegration
When data are non-stationary purely due to unit roots, they can be brought back to stationarity
by linear transformations, for example, by diﬀerencing, as in x
t
− x
t−1
. If x
t
∼ I(1), then by
deﬁnition ∆x
t
∼ I(0). An alternative is to try a linear transformation like y
t
− β
1
x
t
− β
0
, which
induces cointegration when y
t
−β
1
x
t
−β
0
∼ I(0). But unlike diﬀerencing, there is no guarantee that
y
t
−β
1
x
t
−β
0
is I(0) for any value of β, as the discussion in Section 4 demonstrated.
There are many possible tests for cointegration: the most general of them is the multivariate
test based on the vector autoregressive representation (VAR) discussed in Johansen (1988). These
procedures will be described in Part II. Here we only consider tests based on the static and the
dynamic regression model, assuming that x
t
can be treated as weakly exogenous for the parameters
of the conditional model (see e.g., Engle, Hendry and Richard, 1983).
11
As discussed in Section 5,
the condition that there exists a genuine causal link between I(1) series y
t
and x
t
is that the residual
11
Regression methods can be applied to model I(1) variables which are in fact linked (i.e., cointegrated). Most tests
still have conventional distributions, apart from that corresponding to a test for a unit root.
16
−.5 −.25 0 .25 .5 .75 1
.5
1
1.5
2
T=25
−.5 −.25 0 .25 .5 .75 1
1
2
3
4
T=100
−.5 −.25 0 .25 .5 .75 1
2.5
5
7.5
T=400
−.5 −.25 0 .25 .5 .75 1
5
10
15
T=1000
Figure 6: Frequency distributions of stationary autoregression estimators
u
t
∼ I(0), otherwise a ‘nonsense regression’ has been estimated. Therefore, the Engle–Granger
test procedure is based on testing that the residuals u
t
from the static regression model (18) are
stationary, i.e., that ρ < 1 in (2). As discussed in the previous section, the test of the null of a unit
coeﬃcient, using the DF test, implies using a non-standard distribution.
Let ´ u
t
= y
t

´
β
1
x
t

´
β
0
where
´
β is the OLS estimate of the long-run parameter vector β, then
the null hypothesis of the DF test is H
0
: ρ = 1, or equivalently, H
0
: 1 −ρ = 0 in:
´ u
t
= ρ´ u
t−1

t
(33)
or:
∆´ u
t
= (1 −ρ)´ u
t−1

t
.
The test is based on the assumption that ε
t
in (33) is white noise, and if the AR(1) model in (33)
does not deliver white-noise errors, then it has to be augmented by lagged diﬀerences of residuals:
∆´ u
t
= (1 −ρ)´ u
t−1

1
∆´ u
t−1
+ +ψ
m
∆´ u
t−m

t
. (34)
We call the test of H
0
: 1−ρ = 0 in (34) the augmented Dickey–Fuller test (ADF). A drawback of the
DF-type test procedure (see Campos, Ericsson and Hendry, 1996, for a discussion of this drawback)
is that the autoregressive model (33) for ´ u
t
is the equivalent of imposing a common dynamic factor
on the static regression model:
(1 −ρL)y
t
= β
0
(1 −ρ) +β
1
(1 −ρL)x
t

t
. (35)
17
For the DF test to have high power to reject H
0
: 1 − ρ = 0 when it is false, the common-factor
restriction in (35) should correspond to the properties of the data. Empirical evidence has not pro-
duced much support for such common factors, rendering such tests non-optimal. Instead, Kremers,
Ericsson and Dolado (1992) contrast them with a direct test for H
0
: α
2
= 0 in:
∆y
t
= α
0

1
∆x
t

2
(y
t−1
−β
1
x
t−1
−β
0
) +ε
t
, (36)
where the parameters ¦α
0
, α
1
, α
2
, β
0
, β
1
¦ are not constrained by the common-factor restriction in
(35). Unfortunately, the null rejection frequency of their test depends on the values of the ‘nuisance’
parameters α
1
and σ
2
ε
, so Kiviet and Phillips (1992) developed a test which is invariant to these
values. The test reported in the empirical application in the next section is based on this test,
and its distribution is illustrated in Figure 4, panel d. Although non-standard, so its critical values
have been separately tabulated, its distribution is much closer to the Student t-distribution than
the Dickey–Fuller, and correspondingly Banerjee et al. (1993) ﬁnd the power of t
α
2
=0
can be high
relative to the DF test. However, when x
t
is not weakly exogenous (i.e., when not only y
t
to the previous equilibrium error as in (36), but also x
t
does), the test is potentially a poor way of
detecting cointegration. In this case, a multivariate test procedure is needed.
8 An empirical illustration
1990 1995
−.5
0
Gasoline price A
1990 1995
−.75
−.5
−.25
0
Gasoline price B
−1 −.75 −.5 −.25 0
1
2
3 Density of A
−1 −.8 −.6 −.4 −.2 0
1
2
3
4
Density of B
0 5 10
.5
1
Correlogram of A
0 5 10
.5
1
Correlogram of B
Figure 7: Gasoline prices at two locations in (log) levels, their empirical densities and autocorrelo-
grams
In this section, we will apply the concepts and ideas discussed above to a data set consisting of
two weekly gasoline prices (P
a,t
and P
b,t
) at diﬀerent locations over the period 1987 to 1998. The
18
data in levels are graphed in Figure 7 on a log scale, and in (log) diﬀerences in Figure 8. The price
levels exhibit ‘wandering’ behavior, though not very strongly, whereas the diﬀerenced series seem
to ﬂuctuate randomly around a ﬁxed mean of zero, in a typically stationary manner. The bimodal
frequency distribution of the price levels is also typical of non-stationary data, whereas the frequency
distribution of the diﬀerences is much closer to normality, perhaps with a couple of outliers. We also
notice the large autocorrelations of the price levels at long lags, suggesting non-stationarity, and the
lack of such autocorrelations for the diﬀerenced prices, suggesting stationarity (the latter are shown
with lines at ±2SE to clarify the insigniﬁcance of the autocorrelations at longer lags).
1990 1995
−.1
0
.1
.2
Gasoline price A
1990 1995
−.1
0
.1
Gasoline price B
−.1 0 .1 .2
10
20
Density of A
−.1 −.05 0 .05 .1 .15
10
20
Density of B
0 5 10
0
1
Correlogram of A
0 5 10
0
1
Correlogram of B
Figure 8: Gasoline prices at two locations in diﬀerences, their empirical density distribution and
autocorrelogram
We ﬁrst report the estimates from the static regression model:
y
t
= β
0

1
x
t
+u
t
,
and then consider the linear dynamic model:
y
t
= a
0
+a
1
x
t
+a
2
y
t−1
+a
3
x
t−1
+
t
. (38)
Consistent with the empirical data, we will assume that E[∆y
t
] = E[∆x
t
] = 0, i.e., there are no
linear trends in the data. Without changing the basic properties of the model, we can then rewrite
(38) in the equilibrium-correction form:
∆y
t
= a
1
∆x
t
+ (a
2
−1) (y −β
0
−β
1
x)
t−1
+
t
(39)
where a
2
,= 1, and:
β
0
=
a
0
1 −a
2
and β
1
=
a
1
+a
3
1 −a
2
. (40)
19
In formulation (39), the model embodies the lagged equilibrium error (y − β
0
− β
1
x)
t−1
, which
captures departures from the long-run equilibrium as given by the static model. As demonstrated
in (21), the equilibrium error will be a stationary process if (a
2
−1) ,= 0 with a zero mean:
E[y
t
−β
0
−β
1
x
t
] = 0, (41)
whereas if (a
2
−1) = 0, there is no adjustment back to equilibrium and the equilibrium error is a
non-stationary process. The link of cointegration to the existence of a long-run solution is manifest
here, since y
t
− β
0
− β
1
x
t
= u
t
∼I(0) implies a well-behaved equilibrium, whereas when u
t
∼I(1),
no equilibrium exists. In (39), (∆y
t
, ∆x
t
) are I(0) when their corresponding levels are I(1), so with

t
∼ I(0), the equation is ‘balanced’ if and only if (y −β
0
−β
1
x)
t
is I(0) as well.
This type of ‘balancing’ occurs naturally in regression analysis when the model formulation
permits it: we will demonstrate empirically in (43) that one does not need to actually write the
model as in (39) to obtain the beneﬁts. What matters is whether the residuals are uncorrelated or
not.
The estimates of the static regression model over 1987(24)–1998(29) are:
p
a,t
= 0.018
(2.2)
+ 1.01
(67.4)
p
b,t
+u
t
R
2
= 0.89, ´ σ
u
= 0.050, DW = 0.18
(42)
where DW is the Durbin–Watson test statistic for ﬁrst-order autocorrelation, and the ‘t-statistics’
based on (26) are given in parentheses. Although the DW test statistic is small and suggests non-
stationarity, the DF test of u
t
in (42) supports stationarity (DF = −8.21
∗∗
).
Furthermore, the following mis-speciﬁcation tests were calculated:
AR(1–7), F(7, 569) = 524.6 [0.00]
∗∗
ARCH(7), F(7, 562) = 213.2 [0.00]
∗∗
Normality, χ
2
(2) = 22.9 [0.00]
∗∗
The AR(1–m) is a test of residual autocorrelation of order m distributed as F(m, T), i.e. a test of
H
0
: u
t
= ε
t
against H
1
: u
t
= ρ
1
u
t−1
+ +ρ
m
u
t−m

t
. The test of autocorrelated errors of order
1-7 is very large and the null of no autocorrelation is clearly rejected. The ARCH(m) (see Engle,
1982) is a test of autoregressive residual heteroscedasticity of order m distributed as F(m, T − m).
Normality denotes the Doornik and Hansen (1994) test of residual normality, distributed as χ
2
(2). It
is based on the third and the fourth moments around the mean, i.e., it tests for skewness and excess
kurtosis of the residuals. Thus, the normality and homoscedasticity of the residuals are also rejected.
So the standard assumptions underlying the static regression model are clearly violated. In Figure 9,
panel (a), we have graphed the actual and ﬁtted values from the static model (42), in (b) the residuals
´ u
t
, in (c) their correlogram and in (d) the residual histogram compared with the normal distribution.
The residuals show substantial temporal correlation, consistent with a highly autocorrelated process.
This is further conﬁrmed by the correlogram, exhibiting large, though declining, autocorrelations.
This can explain why the DF test rejected the unit-root hypothesis above: though ρ is close to
unity, the large sample size improves the precision of the test and, hence, allows us to reject the
hypothesis. Furthermore, the residuals seem to be symmetrically distributed around the mean, but
are leptokurtic to some extent.
To evaluate the forecasting performance of the model, we have calculated the one-step ahead
forecasts and their (calculated) 95% conﬁdence intervals over the last two years. The outcome
is illustrated in Figure 10a, which shows periods of consistent over- and under- predictions. In
20
1990 1995
−.75
−.5
−.25
0 Actual and fitted values
for A
Actual
Fitted
1990 1995
−2
0
2
4
Residuals
0 5 10
.25
.5
.75
1
Residual correlogram
−4 −2 0 2 4
.2
.4
Residual density
N(0,1)
Figure 9: Actual and ﬁtted values, residuals, their correlogram and histogram, from the static model
particular, at the end of 1997 until the beginning of 1998, predictions were consistently below actual
prices, and thereafter consistently above.
12
Finally, we have recursively calculated the estimated β
0
and β
1
coeﬃcients in (42) from 1993
onwards. Figure 11 shows these recursive graphs. It appears that even if the recursive estimates of
β
0
and β
1
are quite stable, at the end of the period they are not within the conﬁdence band at the
beginning of the recursive sample (and remember the previous footnote). We also report the 1-step
residuals with ±2SE, and the sequence of constancy tests based on Chow (1964) statistics (scaled
by their 1% critical values, so values > 1 reject).
Altogether, most empirical economists would consider this econometric outcome as quite unsat-
isfactory. We will now demonstrate how much the model can be improved by accounting for the
left-out dynamics.
The estimate of the dynamic regression model (38) with two lags is:
p
a,t
= 0.001
(0.3)
+ 1.33
(36.1)
p
a,t−1
− 0.46
(12.5)
p
a,t−2
+ 0.91
(23.9)
p
b,t
− 1.12
(14.8)
p
b,t−1
+ 0.34
(6.4)
p
b,t−2

t
R
2
= 0.99, ´ σ
ε
= 0.018, DW = 2.03
(43)
Compared with the static regression model, we notice that the DW statistic is now close to 2, and
that the residual standard error has decreased from 0.050 to 0.018, so the precision has increased
12
The software assumes the residuals are homoscedastic white-noise when computing conﬁdence intervals, standard
errors, etc., which is ﬂagrantly wrong here.
21
1996 1997 1998
−.75
−.5
−.25
95% confidence bands
Actual
Static model fitted
Static model forecasts
1996 1997 1998
−.75
−.5
−.25
95% confidence bands
Actual
Dynamic model fitted
Dynamic model forecasts
1990 1995
−.1
0
.1
.2
Dynamic model equilibrium error
1990 1995
−.1
0
.1
.2 ∆p
a,t

ecm
t−1

Figure 10: Static and dynamic model forecasts, the equilibrium error, and its relation to ∆p
a,t
approximately 2.5 times. The mis-speciﬁcation tests are:
AR(1–7), F(7, 563) = 1.44 [0.19]
ARCH(7), F(7, 556) = 10.6 [0.00]
∗∗
Normality, χ
2
(2) = 79.8 [0.00]
∗∗
We note that residual autocorrelation is no longer a problem, but there is still evidence of the
residuals being heteroscedastic and non-normal. Figure 12 show the graphs of actual and ﬁtted,
residuals, residual correlogram, and the residual histogram compared to the normal distribution.
Fitted and actual values are now so close that it is no longer possible to visually distinguish between
them. Residuals exhibit no temporal correlation. However, with the increased precision (the smaller
residual standard error), we can now recognize several ‘outliers’ in the data. This is also evident
from the histogram exhibiting quite long tails, resulting in the rejection of normality. The rejection
of homoscedastic residual variance seems to be explained by a larger variance in the ﬁrst part of the
sample, prevailing approximately till the end of 1992, probably due to the Gulf War.
As for the static regression model, the one-step ahead prediction errors with their 95% conﬁdence
intervals have been graphed in Figure 10b. It appears that there is no systematic under- or over-
prediction in the dynamic model. Also the prediction intervals are much narrower, reﬂecting the
large increase in precision as a result of the smaller residual standard error in this model.
It is now possible to derive the static long-run solution as given by (39) and (40):
p
a,t
= 0.008
(0.3)
+ 0.99
(22.8)
p
b,t
t
ur
= 8.25
∗∗
(44)
22
1990 1995
−.1
−.05
0
.05
.1
Recursively−computed intercept
with ±2SE, static model
1990 1995
.8
.9
1
1.1
Recursively−computed slope
with ±2SE, static model
1990 1995
−.1
0
.1
.2
1−step residuals with±2SE
1990 1995
.25
.5
.75
1
Recursively−computed constancy test
Figure 11: Recursively-calculated coeﬃcients of β
0
and β
1
in (41) with 1-step residuals and Chow
tests
Note that the coeﬃcient estimates of β
0
and β
1
are almost identical to the static regression model,
illustrating the fact that the OLS estimator was unbiased. However, the correctly-calculated standard
errors of estimates produce much smaller t-values, consistent with the downward bias of SE[
´
β] in
the static regression model, revealing an insigniﬁcant intercept. The graph of the equilibrium error
u
t
(shown in Figure 10c) is essentially the log of the relative price, p
a,t
− p
b,t
. Note that u
t
has a
zero mean, consistent with (41), and that it is strongly autocorrelated, consistent with (21). From
the coeﬃcient estimate of the ecm
t−1
below in (45), we ﬁnd that the autocorrelation coeﬃcient
(1 − α
2
) in (21) corresponds to 0.86, which is a fairly high autocorrelation. In the static model,
this autocorrelation was left in the residuals, whereas in the dynamic model, we have explained it
by short-run adjustment to current and lagged changes in the two gasoline prices. Nevertheless,
the values of the DF test on the static residuals, and the unit-root test in the dynamic model (t
ur
in (44)) are closely similar here, even though the test for two common factors in (43) rejects (one
common factor is accepted).
Finally, we report the estimates of the model reformulated in equilibrium-correction form (39),
suppressing the constant of zero due to the equilibrium-correction term being mean adjusted:
∆p
a,t
= 0.46
(12.5)
∆p
a,t−1
+ 0.91
(24.3)
∆p
b,t
− 0.34
(6.5)
∆p
b,t−1
− 0.13
(8.3)
ecm
t−1

t
R
2
= 0.69, ´ σ
ε
= 0.018, DW = 2.05
(45)
Notice that the corresponding estimated coeﬃcients, the residual standard error, and DW are
identical with the estimates in (43), demonstrating that the two models are statistically equivalent,
though one is formulated in non-stationary and the other in stationary variables. The coeﬃcient
23
1990 1995
−.75
−.5
−.25
0
Actual
Fitted
1990 1995
−2.5
0
2.5
Residual
0 5 10
−.5
0
.5
1
Correlogram
−4 −2 0 2 4
.2
.4
Density
N(0,1)
Figure 12: Actual and ﬁtted values, residuals, their correlogram and histogram for the dynamic
model
of −0.13 on ecm
t−1
suggests moderate adjustment, with 13% of any disequilibrium in the relative
prices being removed each week.
13
Figure 10d shows the relation of ∆p
a,t
to ecm
t−1
.
However, R
2
is lower in (45) than in (42) and (43), demonstrating the hazards of interpreting
R
2
as a measure of goodness of ﬁt. When we calculate R
2
= Σ(y
t
− ´ y
t
)
2
/Σ(y
t
− y)
2
, we essentially
compare how well our model can predict y
t
compared to a straight line. When y
t
is trending, any
other trending variable can do better than a straight line, which explains the high R
2
often obtained
in regressions with non-stationary variables. Hence, as already discussed in Section 5, the sum of
squares Σ(y
t
−y)
2
is not an appropriate measure of the variation of a trending variable. In contrast,
R
2
= Σ(∆y
t

´
∆y
t
)
2
/Σ(∆y
t
− ∆y)
2
from (45) measures the improvement in model ﬁt compared
to a random walk as a reasonable measure of how good our model is (although that R
2
would also
change if the dependent variable became ecm
t
in another statistically-equivalent version of the same
model).
Finally, we report the recursively-calculated coeﬃcients of the model parameters in (45) in Figure
13. The estimated coeﬃcients are reasonably stable over time, although a few of the recursive estim-
ates fall outside the conﬁdence bands (especially around the Gulf War). Such non-constancies reveal
remaining non-stationarities in the two gasoline prices, but resolving this issue would necessitate
another paper.....
13
The equilibrium-correction error ecm
t−1
does not inﬂuence ∆p
b,t
in a bivariate system, so that aspect of weak
exogeneity is not rejected.
24
1990 1995
.4
.6
.8
∆p
a,t−1

1990 1995
.25
.5
.75
1
1.25
∆p
b,t

1990 1995
−.75
−.5
−.25
0
.25
∆p
b,t−1

1990 1995
−.3
−.2
−.1
0
ecm
t−1

Figure 13: Recursively-calculated parameter estimates of the ECM model
9 Conclusion
Although ‘classical’ econometric theory generally assumed stationary data, particularly constant
means and variances across time periods, empirical evidence is strongly against the validity of that
assumption. Nevertheless, stationarity is an important basis for empirical modeling, and inference
when the stationarity assumption is incorrect can induce serious mistakes. To develop a more
relevant basis, we considered recent developments in modeling non-stationary data, focusing on
autoregressive processes with unit roots. We showed that these processes were non-stationary, but
could be transformed back to stationarity by diﬀerencing and cointegration transformations, where
the latter comprised linear combinations of the variables that did not have unit roots.
We investigated the comparative properties of stationary and non-stationary processes, reviewed
the historical development of modeling non-stationarity and presented a re-run of a famous Monte
Carlo simulation study of the dangers of ignoring non-stationarity in static regression analysis. Next,
we described how to test for unit roots in scalar autoregressions, then extended the approach to tests
for cointegration. Finally, an extensive empirical illustration using two gasoline prices implemented
the tools described in the preceding analysis.
Unit-root non-stationarity seems widespread in economic time series, and some theoretical models
entail unit roots. Links between variables will then ‘spread’ such non-stationarities throughout the
economy. Thus, we believe it is sensible empirical practice to assume unit roots in (log) levels until
that is rejected by well-based evidence. Cointegrated relations and diﬀerenced data both help model
unit roots, and can be related in equilibrium-correction equations, as we illustrated. For modeling
purposes, a unit-root process may also be considered as a statistical approximation when serial
25
correlation is high. Monte Carlo studies have demonstrated that treating near-unit roots as unit
roots in situations where the unit-root hypothesis is only approximately correct makes statistical
inference more reliable than otherwise.
Unfortunately, other sources of non-stationarity may remain, such as changes in parameters (par-
ticularly shifts in the means of equilibrium errors and growth rates) or data distributions, so careful
empirical evaluation of ﬁtted equations remains essential. We reiterate the importance of having
white-noise residuals, preferably homoscedastic, to avoid mis-leading inferences. This emphasizes the
advantages of accounting for the dynamic properties of the data in equilibrium-correction equations,
which not only results in improved precision from lower residual variances, but delivers empirical
estimates of adjustment parameters.
Part II of our attempt to explain cointegration analysis will address system methods. Since
cointegration inherently links several variables, multivariate analysis is natural, and recent develop-
ments have focused on this approach. Important new insights result, but new modeling decisions
also have to made in practice. Fortunately, there is excellent software available for implementing the
methods discussed in Johansen (1995a), including CATS in RATS (see Hansen and Juselius, 1995)
and PcFiml (see Doornik and Hendry, 1997), and we will address the application of such tools to
the gasoline price data considered above.
References
Attﬁeld, C. L. F., Demery, D., and Duck, N. W. (1995). Estimating the UK demand for money
function: A test of two approaches. Mimeo, Economics department, University of Bristol.
Baba, Y., Hendry, D. F., and Starr, R. M. (1992). The demand for M1 in the U.S.A., 1960–1988.
Review of Economic Studies, 59, 25–61.
Banerjee, A., Dolado, J. J., Galbraith, J. W., and Hendry, D. F. (1993). Co-integration, Error
Correction and the Econometric Analysis of Non-Stationary Data. Oxford: Oxford University
Press.
Banerjee, A., and Hendry, D. F. (1992). Testing integration and cointegration: An overview. Oxford
Bulletin of Economics and Statistics, 54, 225–255.
Campos, J., Ericsson, N. R., and Hendry, D. F. (1996). Cointegration tests in the presence of
structural breaks. Journal of Econometrics, 70, 187–220.
Chan, N. H., and Wei, C. Z. (1988). Limiting distributions of least squares estimates of unstable
autoregressive processes. Annals of Statistics, 16, 367–401.
Chow, G. C. (1964). A comparison of alternative estimators for simultaneous equations. Economet-
rica, 32, 532–553.
Davidson, J. E. H., Hendry, D. F., Srba, F., and Yeo, J. S. (1978). Econometric modelling of the
aggregate time-series relationship between consumers’ expenditure and income in the United
Kingdom. Economic Journal, 88, 661–692. Reprinted in Hendry, D. F. (1993), Econometrics:
Alchemy or Science? Oxford: Blackwell Publishers.
Dickey, D. A., and Fuller, W. A. (1979). Distribution of the estimators for autoregressive time series
with a unit root. Journal of the American Statistical Association, 74, 427–431.
Dickey, D. A., and Fuller, W. A. (1981). Likelihood ratio statistics for autoregressive time series
with a unit root. Econometrica, 49, 1057–1072.
26
Doornik, J. A., and Hansen, H. (1994). A practical test for univariate and multivariate normality.
Discussion paper, Nuﬃeld College.
Doornik, J. A., and Hendry, D. F. (1997). Modelling Dynamic Systems using PcFiml 9 for Windows.
London: Timberlake Consultants Press.
Doornik, J. A., and Hendry, D. F. (1998). Monte Carlo simulation using PcNaive for Windows.
Unpublished typescript, Nuﬃeld College, University of Oxford.
Engle, R. F. (1982). Autoregressive conditional heteroscedasticity, with estimates of the variance of
United Kingdom inﬂations. Econometrica, 50, 987–1007.
Engle, R. F., and Granger, C. W. J. (1987). Cointegration and error correction: Representation,
estimation and testing. Econometrica, 55, 251–276.
Engle, R. F., Hendry, D. F., and Richard, J.-F. (1983). Exogeneity. Econometrica, 51, 277–304.
Reprinted in Hendry, D. F., Econometrics: Alchemy or Science? Oxford: Blackwell Publishers,
1993; and in Ericsson, N. R. and Irons, J. S. (eds.) Testing Exogeneity, Oxford: Oxford
University Press, 1994.
Friedman, M., and Schwartz, A. J. (1982). Monetary Trends in the United States and the United
Kingdom: Their Relation to Income, Prices, and Interest Rates, 1867–1975. Chicago: Uni-
versity of Chicago Press.
Granger, C. W. J. (1981). Some properties of time series data and their use in econometric model
speciﬁcation. Journal of Econometrics, 16, 121–130.
Granger, C. W. J., and Newbold, P. (1974). Spurious regressions in econometrics. Journal of
Econometrics, 2, 111–120.
Hall, P., and Heyde, C. C. (1980). Martingale Limit Theory and its Applications. London: Academic
Press.
Hansen, H., and Juselius, K. (1995). CATS in RATS: Cointegration analysis of time series. Discussion
paper, Estima, Evanston, IL.
Hendry, D. F. (1980). Econometrics: Alchemy or science?. Economica, 47, 387–406. Reprinted in
Hendry, D. F. (1993), Econometrics: Alchemy or Science? Oxford: Blackwell Publishers.
Hendry, D. F. (1995). Dynamic Econometrics. Oxford: Oxford University Press.
Hendry, D. F., and Anderson, G. J. (1977). Testing dynamic speciﬁcation in small simultaneous
systems: An application to a model of building society behaviour in the United Kingdom. In
Intriligator, M. D. (ed.), Frontiers in Quantitative Economics, Vol. 3, pp. 361–383. Amster-
dam: North Holland Publishing Company. Reprinted in Hendry, D. F. (1993), Econometrics:
Alchemy or Science? Oxford: Blackwell Publishers.
Hendry, D. F., and Mizon, G. E. (1978). Serial correlation as a convenient simpliﬁcation, not a
nuisance: A comment on a study of the demand for money by the Bank of England. Economic
Journal, 88, 549–563. Reprinted in Hendry, D. F. (1993), Econometrics: Alchemy or Science?
Oxford: Blackwell Publishers.
Hendry, D. F., and Morgan, M. S. (1995). The Foundations of Econometric Analysis. Cambridge:
Cambridge University Press.
Johansen, S. (1988). Statistical analysis of cointegration vectors. Journal of Economic Dynamics
and Control, 12, 231–254.
Johansen, S. (1991). Estimation and hypothesis testing of cointegration vectors in Gaussian vector
autoregressive models. Econometrica, 59, 1551–1580.
27
Johansen, S. (1992a). Determination of cointegration rank in the presence of a linear trend. Oxford
Bulletin of Economics and Statistics, 54, 383–398.
Johansen, S. (1992b). A representation of vector autoregressive processes integrated of order 2.
Econometric Theory, 8, 188–202.
Johansen, S. (1995a). Likelihood-based Inference in Cointegrated Vector Autoregressive Models. Ox-
ford: Oxford University Press.
Johansen, S. (1995b). Likelihood based Inference on Cointegration in the Vector Autoregressive
Model. Oxford: Oxford University Press.
Johansen, S., and Juselius, K. (1990). Maximum likelihood estimation and inference on cointegration
– With application to the demand for money. Oxford Bulletin of Economics and Statistics,
52, 169–210.
Johansen, S., and Juselius, K. (1992). Testing structural hypotheses in a multivariate cointegration
analysis of the PPP and the UIP for UK. Journal of Econometrics, 53, 211–244.
Kiviet, J. F., and Phillips, G. D. A. (1992). Exact similar tests for unit roots and cointegration.
Oxford Bulletin of Economics and Statistics, 54, 349–367.
Kremers, J. J. M., Ericsson, N. R., and Dolado, J. J. (1992). The power of cointegration tests.
Oxford Bulletin of Economics and Statistics, 54, 325–348.
Morgan, M. S. (1990). The History of Econometric Ideas. Cambridge: Cambridge University Press.
Nelson, C. R., and Plosser, C. I. (1982). Trends and random walks in macroeconomic time series:
some evidence and implications. Journal of Monetary Economics, 10, 139–162.
Park, J. Y., and Phillips, P. C. B. (1988). Statistical inference in regressions with integrated pro-
cesses. part 1. Econometric Theory, 4, 468–497.
Park, J. Y., and Phillips, P. C. B. (1989). Statistical inference in regressions with integrated pro-
cesses. part 2. Econometric Theory, 5, 95–131.
Phillips, P. C. B. (1986). Understanding spurious regressions in econometrics. Journal of Econo-
metrics, 33, 311–340.
Phillips, P. C. B. (1987a). Time series regression with a unit root. Econometrica, 55, 277–301.
Phillips, P. C. B. (1987b). Towards a uniﬁed asymptotic theory for autoregression. Biometrika, 74,
535–547.
Phillips, P. C. B. (1988). Regression theory for near-integrated time series. Econometrica, 56,
1021–1043.
Phillips, P. C. B., and Perron, P. (1988). Testing for a unit root in time series regression. Biometrika,
75, 335–346.
Qin, D. (1993). The Formation of Econometrics: A Historical Perspective. Oxford: Clarendon
Press.
Sargan, J. D. (1964). Wages and prices in the United Kingdom: A study in econometric methodology
(with discussion). In Hart, P. E., Mills, G., and Whitaker, J. K. (eds.), Econometric Analysis
for National Economic Planning, Vol. 16 of Colston Papers, pp. 25–63. London: Butterworth
Co. Reprinted as pp. 275–314 in Hendry D. F. and Wallis K. F. (eds.) (1984). Econometrics
and Quantitative Economics. Oxford: Basil Blackwell, and as pp. 124–169 in Sargan J. D.
(1988), Contributions to Econometrics, Vol. 1, Cambridge: Cambridge University Press.
Spanos, A. (1986). Statistical Foundations of Econometric Modelling. Cambridge: Cambridge Uni-
versity Press.
28
Stock, J. H. (1987). Asymptotic properties of least squares estimators of cointegrating vectors.
Econometrica, 55, 1035–1056.
Student (1908). On the probable error of the mean. Biometrika, 6, 1–25.
Working, H. (1934). A random diﬀerence series for use in the analysis of time series. Journal of the
American Statistical Association, 29, 11–24.
Yule, G. U. (1926). Why do we sometimes get nonsense-correlations between time-series? A study
in sampling and the nature of time series (with discussion). Journal of the Royal Statistical
Society, 89, 1–64.
29

(a)
12.5 10 7.5 5 1875 1900 1925 1950 1975 2000 1.5
UK nominal money UK price level

(b)
9 8 7
UK real money UK real output

(c)

1875

1900

1925

1950

1975

2000

(d)
US real output US real money

.15 .1 .05

UK short−term interest rate UK long−term interest rate

1

1875 .15 .1

1900

1925

1950

1975

2000 .2

1960

1970
UK annual inflation US annual inflation

1980

1990

(e)
US short−term interest rate US long−term interest rate

(f)

.1 .05 0 1950 1960 1970 1980 1990 1950 1960 1970 1980 1990

Figure 1: Some ‘representative’ time series the relationships between them: all quiet till 1929, the two variables diverge markedly till the early 1950s, then rise together with considerable variation. Nor is the non-stationarity problem speciﬁc to the UK: panels d and e show the comparative graphs to b and c for the USA, using post-war quarterly data (from Baba, Hendry and Starr, 1992). Again, there is considerable evidence of change, although the last panel f comparing UK and US annual inﬂation rates suggests the UK may exhibit greater instability. It is hard to imagine any ‘revamping’ of the statistical assumptions such that these outcomes could be construed as drawings from stationary processes.2 Intermittent episodes of forecast failure (a signiﬁcant deterioration in forecast performance relative to the anticipated outcome) conﬁrm that economic data are not stationary: even poor models of stationary data would forecast on average as accurately as they ﬁtted, yet that manifestly does not occur empirically. The practical problem facing econometricians is not a plethora of congruent models from which to choose, but to ﬁnd any relationships that survive long enough to be useful. It seems clear that stationarity assumptions must be jettisoned for most observable economic time series. Four issues immediately arise: 1. how important is the assumption of stationarity for modeling and inference? 2. what are the eﬀects of incorrectly assuming it? 3. what are the sources of non-stationarity?
2 It is sometimes argued that economic time series could be stationary around a deterministic trend, and we will comment on that hypothesis later.

2

but to develop the analysis. technology involves the persistence of acquired knowledge. often inducing structural breaks in time series.4. including the venerable linear trend. as we will show. legislative changes. For example. but it is far from the only one. building on the famous study in Yule (1926). ensures secular trends in many time series. and deﬁne the properties of a stationary and a non-stationary process (Section 3). so that the present level of technology is the cumulation of past discoveries and innovations. Section 5 uses simulation experiments to look at the consequences of data being non-stationary in regression models. Finally.3 Such processes can be interpreted as allowing a diﬀerent ‘trend’ at every point in time. can empirical analyses be transformed so stationarity becomes a valid assumption? Essentially. assuming constant means and variances when that is false can induce serious statistical mistakes. non-stationarity can be due to evolution of the economy. some forms of non-stationarity can be eliminated by transformations. 2. Our focus here will be on a type of stochastic non-stationarity induced by persistent cumulation of past eﬀects. Only intuitive explanations for these answers can be oﬀered at this stage of our paper. Other variables related to 3 Stochastic means the presence of a random variable. 4. we next review the history of regression analyses with trending data (Section 4) and consider the possibilities of obtaining stationarity by transformation (called cointegration). and political turmoil inter alia. As embryology often yields insight into evolution. perhaps resulting from technological progress. 3 . posing diﬃcult problems for empirical modeling. Section 9 concludes. as well as the validity of the transformation in Section 4 (cointegration tests: Section 7). observations come from diﬀerent distributions over time. called unit-root processes (an explanation for this terminology is provided below). Legislative change is one obvious source of non-stationarity. 2 Addressing non-stationarity Non-stationarity seems a natural feature of economic life. when data means and variances are non-constant. we empirically illustrate the concepts and ideas for a data set consisting of gasoline prices in two major locations (Section 8). and much of our paper concerns an important case where that is indeed feasible. depending on the source of non-stationarity’. We expand on all four of these issues below. Such trends need to be incorporated into statistical analyses. Economic growth. so are said to have stochastic trends. but roughly: 1. We then brieﬂy review tests to determine the presence of non-stationarity in the class noted in Section 5 (called univariate unit-root processes: Section 6). and ‘sometimes. 3. ‘potentially hazardous’. Economic variables depending closely on technological progress are therefore likely to have a stochastic trend. There are many plausible reasons why economic data may contain stochastic trends. which could be done in many ways. we ﬁrst discuss some necessary econometric concepts based on a simple model (in Section 2). technological change. The impact of structural changes in the world oil market is another example of non-stationarity. ‘many and varied’. Figure 2 provides an artiﬁcial example: ‘joining’ the successive observations as shown produces a ‘jumpy’ trend as compared to the (best-ﬁtting) linear trend. as Figure 1 illustrated. the answers are ‘very’.

so the linkages in economies suggest that the levels of many variables will be non-stationary. a variable exhibiting a shift in its mean is a non-stationary process. and transmit it to other variables in turn: nominal wealth and exports spring to mind. (1) To make the example more realistic. as is a variable with a heteroscedastic variance over time. and therefore income and expenditure. To introduce the basic econometric concepts. T. . For example.10 8 6 4 2 1 2 3 4 5 6 7 8 9 10 Figure 2: An artiﬁcial variable with a stochastic trend the level of any variable with a stochastic trend will ‘inherit’ that non-stationarity. the current value of the variable ut is aﬀected by its value in the immediately preceding period (ut−1 ) with a coeﬃcient ρ. (2) That is. and discuss its implications for empirical modeling. one which violates the stationarity requirement. we consider a simple regression model for a variable yt containing a ﬁxed (or deterministic) linear trend with slope β generated from an initial value y0 by: yt = y0 + βt + ut for t = 1. sharing a set of common stochastic trends. so its means and variances are non-constant over time. The stochastic ‘shock’ εt is distributed as IN[0. and by a stochastic ‘shock’ εt . by deﬁnition. A non-stationary process is. and so employment. We will focus here on the non-stationarity caused by stochastic trends... wages etc.. Similar consequences follow for every source of stochastic trends. σ 2 ]. the error term ut is allowed to be a ﬁrst-order autoregressive process: ut = ρut−1 + εt . We discuss the meaning of the autoregressive parameter ρ below. denoting an ε 4 .

(6) εt + y0 1 − ρL From (6). I(2)). and normality. While there may appear to be many names for the same notion. longer lags in (2) preclude ut being a random walk. and ut ∼ I(0) when |ρ| < 1. we will use the notation ut for an autocorrelated process. the larger the value of ρ. extensions yield important distinctions: for example. normal (N) distribution with a mean of zero (E [εt ] = 0) and a variance V [εt ] = σ 2 : ε since these are constant parameters. when |ρ| < 1. 1 − ρ2 (5) Hence. but that the eﬀects of previous disturbances decline with time because |ρ| < 1. we will use lower-case letters to indicate logarithmic scale. In that case.e. the variance of ut becomes indeterminate and ut becomes a random walk. we notice that ρ = 1 is equivalent to the summation of the errors. summation corresponds to integration. ut = εt +εt−1 +εt−2 +· · · . and εt for a white-noise process.. 1995. The diﬀerence between a linear stochastic trend and i=1 a deterministic trend is that the increments of a stochastic trend are random. identical distributions. so we ﬁrst consider its properties. the ﬁrst two are clearly the most important. and why an autoregressive error imposes a common-factor dynamics on a static 5 . Throughout the paper. The process ut in (2) is not independent. Then. we say that ut has the ‘stochastic trend’ Σt εi . here of ﬁrst order: we use the shorthand notation ut ∼ I(1) when ρ = 1. However. The assumption of only ﬁrst-order autocorrelation in ut . chapter 4) such that Lxt = xt−1 . (2) is called a unit-root process. so have several unit roots. now think of (4) as a process directly determining ut – ignoring our derivation from (2) – and consider what happens when ρ = 1. (2) can be written as ut = ρLut + εt or: ut = εt . In continuous time. From (4). so xt = log(Xt ). From (4).g. Consequently. and processes can be integrated of order 2 (i. Hendry. Interpreted as a polynomial in L. so such processes are also called integrated. the term 1/(1 − ρL) in (3) can be expanded as (1 + ρL + ρ2 L2 + · · · ). which has a root of 1/ρ: when ρ = 1. an identical distribution holds at every point in time. In the following. 1 − ρL (3) When |ρ| < 1. is often called a unit-root process. by substituting (2) into (1) we get: yt = βt + and by multiplying through the factor (1 − ρL): (1 − ρL)yt = (1 − ρL)βt + (1 − ρL)y0 + εt . and all arguments generalize to higher-order autoregressions. When ρ = 1. (4) Expression (4) can also be derived after repeated substitution in (2). Returning to the trend regression example. it is easy to see why the non-stationary process which results when ρ = 1. A process such as {εt } is often called a normal ‘white-noise’ process. whereas those of a deterministic trend are constant over time as Figure 2 illustrated. Hence: ut = εt + ρεt−1 + ρ2 εt−2 + · · · . (3) has a factor of 1 − ρL.. as shown in (2). Of the three desirable properties of independence. is for notational simplicity. the larger the variance of ut . Dynamic processes are most easily studied using the lag operator L (see e. so each disturbance persists indeﬁnitely and has a permanent eﬀect on ut .independent (I). we can derive the properties of ut as: E [ut ] = 0 and V [ut ] = σ2 ε . It appears that ut is the sum of all previous disturbances (shocks) εt−i .

From Case 1. where ∆yt = yt −yt−1 . 6 . |ρ| < 1 and β = 0 delivers yt = ρyt−1 + (1 − ρ)y0 + εt . other asset prices..4 Case 4.. We will now provide some examples of economic variables that can be appropriately described by the above simple models. so it describes a linear diﬀerence equation with a unit coeﬃcient. yt contains no linear trend. • in the diﬀerenced model (ρ = 1). It follows from (7) that ∆yt = β + εt . and prices of widely-traded commodities. . . b2 . since although the coeﬃcient of t in (7) is zero. Hence: • in the static regression model (1). Similar arguments could apply to exchange rates. 3 Properties of a non-stationary and a stationary process Unit-root processes can also arise as a consequence of plausible economic behavior. • in the dynamic regression model (8). which is the usual stationary autoregressive model with a constant term. the constant term is solely measuring the growth rate. the constant term is essentially accounting for the unit of measurement of yt . Note that E[∆yt ] = β = 0 is equivalent to yt having a linear trend. As an example. we will discuss the possibility of a unit root in the long-term interest rate.regression model (see e.e. it follows immediately that ∆yt = εt . The interpretation of the coeﬃcients b1 . i.. b2 is not an estimate of the trend in yt . a ‘trend-stationary’ model. the ‘kick-oﬀ’ value of the y series. the root of the lag polynomial is unity. because of the thought experiment that output would then continue to grow if we all ceased working. and b0 must be done with care: for example.e. |ρ| < 1 and β = 0 gives us (8). β. Rewriting (6) using Lxt = xt−1 . . instead β = b2 /(1 − ρ) is the trend in the process. ρ = 1 and β = 0. 4 We doubt that GNP (say) has a deterministic trend. . two of which correspond to non-stationary (unit-root) models.. 1978). Case 3. the constant term is a weighted average of the growth rate β and the initial value y0 . This model is popularly called a ‘random walk with drift’.. When ρ = 1 in (6). Hendry and Mizon. just as ut cumulated past εt in (4). for t = 1. and the other two to stationary models: Case 1. the coeﬃcient of yt−1 is unity so it ‘integrates’ the ‘intercept’ β. ρ = 1 and β = 0. we get: yt = ρyt−1 + β(1 − ρ)t + ρβ + (1 − ρ)y0 + εt . (9) (8) We will now consider four diﬀerent cases.. T . i.. such as gasoline. (7) and it appears that the ‘static’ regression model (1) with autocorrelated residuals is equivalent to the following dynamic model with white-noise residuals: yt = b1 yt−1 + b2 t + b0 + εt where b1 = ρ b2 = β(1 − ρ) b0 = ρβ + (1 − ρ)y0 . Case 2. This is called a pure random walk model: since E[∆yt ] = 0.g.

and as a dynamic relation.1 A non-stationary process Equation (10) shows that the whole of Rl.t−1 and t inﬂuence Rl.t−1 . in the next period.004. and hence. so we anticipate that the expected value of the change in the long-term interest rate at time t − 1. ε which need not be normal).994 with an estimated standard error of 0.05 0 1 5 10 15 1960 1970 1980 1990 1 0 −1 Change in the long−term interest rate 1950 3 2 1 −1. to compensate lenders for tying up their money – one could create a money machine.5 Correlogram Correlogram . Thus. this translates into: Rl. 5 Empirically. the whole of Rl. the estimated coeﬃcient of Rl. As a model.15 .5 1 1. the change should be small on a risk-adjusted basis after transactions costs).5 0 0 5 10 0 5 10 Figure 3: Monthly long-term interest rates in the USA. we ﬁrst need to discuss the statistical properties of stationary and non-stationary processes. and Rl > Rs (the short-term rate) – as usually holds. 7 .5 1 1960 1970 1980 1990 Density Density −1 −. is zero.t . Such a scenario of boundless proﬁt at low risk seems unlikely.5 Long−term interest rate 15 10 5 1950 .. 1950–1993 3.1 .2 .t = Rl. and past for the monthly data over 1950(1)–1993(12) on 20-year bond rates in the USA shown in Figure 3. Just predict the forthcoming change in Rl . is a unit-root process. and borrow at Rs to buy bonds if you expect a fall in Rl (a rise in bond prices) or sell short if Rl is likely to rise. Et−1 [∆Rl. i.t+1 and so on.5 0 . To discuss the implications for empirical modeling of having unit roots in the data. σ 2 ] process (where D denotes the relevant distribution.e. given the relevant information set It−1 . the eﬀect of t persists indeﬁnitely.t inﬂuences Rl. in levels and diﬀerences.t−1 in (10) is 0.t |It−1 ] = 0 (more generally.If changes to long-term interest rates (Rl ) were predictable. The model in (10) has a unit coeﬃcient on Rl.t−1 + t (10) where Et−1 [ t |It−1 ] = 0 and t is an ID[0.

t |Rl. violating the constant-variance assumption of a stationary process. one can see that: E [Rl. 1] in large samples: see Hendry. Finally. a feature of Rl.0 is constant and equal to Rl. panel e.0 . . In practical applications.t . To summarize. Figure 3. time 0 corresponds to the ﬁrst observation in the sample. the conditional model (12) is more relevant as a description of the sample variation.0 as a ﬁxed number. We will comment on the properties of ∆Rl. but this variance is non-constant since it changes with t = 1.0 (12) where the initial value Rl.0 )] = σ 2 s and so: s corr2 [Rl.t |Rl.t − Rl. taking Rl. 8 .t |Rl.. as illustrated in Figure 3. panels a and b). The theoretical mean of the conditional process Rl. and successive observations are highly interdependent. contains all information of the past behavior of the long-term interest rate up to time 0. Rl. and shows that {Rl. Rl.t = t t + t−1 + ···+ 1 + 0 + −1 ··· (11) + t−1 + ···+ 1 + Rl..t−s ] = E [(Rl.0 and that: V [Rl. . From (12). Therefore.t ] = σ 2 t.s − Rl. . and is. and so does not look even approximately normal. T .0 .t below.0 } has a ﬁnite variance. we will from the outset discuss stationarity either for a diﬀerenced variable {∆yt } or for the IID errors {εt }. all the serial correlations for a random-walk process are close to unity. t (16) (15) Consequently.t = or alternatively: Rl. (14) (13) Further. most economic time series are non-stationary. and in fact. panel c shows the histogram of approximately 80 quarterly observations for which the observed distribution is bimodal.errors accumulate with no ‘depreciation’. even if {Rl. Rl. t = 1. As argued above. when t is large.. the covariance between drawings t − s periods apart is: C [Rl.t .0 = 0 + −1 . Equation (11) shows that theoretically.t } is the sum of a large number of errors it will not be approximately normally distributed. .t as illustrated in Figure 3. at least in comparison to its ﬁrst diﬀerence. the sample mean divided by 3T is distributed as N[0.0 ) (Rl.t ] = Rl. 3. In empirical studies. and at best become stationary only after diﬀerencing. when t > s. 1995.. has a diﬀerent variance.t−s ] = 1 − . Cumulating random errors will make them smooth. so an equivalent formulation of (10) is: Rl. the unit-root assumption implies an ever-increasing variance to the time series (around a ﬁxed mean). However. tσ 2 . induces properties like those of economic variables. chapter 3). T. the variance of a unit-root process increases over time.. perhaps not so easily seen. This is because each observation. . as ﬁrst discussed by Working (1934) (so Rl.t should be smooth. the theoretical mean and the empirical sample mean Rl are not even √ approximately equal when data are non-stationary (surprisingly.2 A stationary process We now turn to the properties of a stationary process.

so the variance of xt increases over time. making it non-stationary. and Hendry and Morgan (1995). because ∆Rl. the data take the form ∆xt = εt (Case 2. and that the density distribution is approximately normal except for an outlier corresponding to the deregulation of capital movements in 1983 (panel d). However. Rather.g. or more precisely. but such that the inﬂuence of past shocks dies out. population growth). we brieﬂy review the evolution of the concepts and tools underpinning the analysis of non-stationarity in economics.6 As an example of a stationary process take the change in the long-term interest rate from (10): ∆Rl. (17) It is a stationary process with mean E[∆Rl. A detailed history of econometrics has also developed over the past decade and full coverage is provided in Morgan (1990). This is illustrated by the graph of the change in the long-term bond rate in Figure 3. in Section 3) where εt is IID.t ] = 0. He thought that they were not the result of both variables being related to some third variable (e.t − 0)2 ] = σ 2 .g. the variance would not be constant. and E[(∆Rl. which are extremely high correlations often found between variables for which there is no ready causal explanation (such as birth rates of humans and the number of storks’ nests in Stockholm). Otherwise. σ 2 . and E[(∆yt − µ)(∆yt−s − µ)] = γ(s) ∀s. as demonstrated in (2). stationarity holds only asymptotically (see e. 6 When the moments depend on the initial conditions of the process. and contains the linear stochastic trend Σεi . An IID process is the simplest example of a stationary process. 4 Regression with trending variables: a historical review One might easily get the impression that the unit-root literature is a recent phenomenon. We note that the autocorrelations have more or less disappeared (panel f). E[(∆yt − µ)2 ] = σ 2 . is quite recent. starting from the initial level x0 at time 0. C] random second diﬀerences. 9 . the level of xt contains information about all permanent disturbances that have aﬀected the variable.. the development of ‘unit-root econometrics’ that aims to address this problem in a more constructive way. although its stochastic trend depends on cumulating errors that are independent of those entering xt . Qin (1993). Udny Yule analyzed the hazards of regressing a trending variable on another unrelated trending variable – the so-called ‘nonsense regression’ problem.A variable ∆yt is weakly stationary when its ﬁrst two moments are constant over time. and γ(s) are ﬁnite and independent of t. E[(∆Rl. Yule (1926) was the ﬁrst to formally analyze such ‘nonsense correlations’.t − 0)(∆Rl. This is clearly not the case: already in 1926. when E[∆yt ] = µ. However. σ 2 . Here. Spanos. as demonstrated by (4). a stationary process can be autocorrelated. B] random ﬁrst diﬀerences. commencing with the problem of ‘nonsense correlations’. namely. he categorized time series according to their serial-correlation properties.t = t where t ∼ ID 0. 1986). panel b.t = εt is assumed to be an IID process. and investigated how their crosscorrelation coeﬃcient rxy behaved when two unconnected series x and y had: A] random levels. where µ. Therefore. Since the value of xt depends on all past errors with equal weights. Similarly for the yt series. but we ignore that complication here. in case B.t−s − ε 0)] = 0 ∀s. how highly correlated successive values were with each other. For example. the eﬀects of distant shocks persist.

and I(2) as above). Granger and Newbold (1974) highlighted that a good ﬁt with signiﬁcant serial correlation in the residuals was a symptom associated with nonsense regressions. and so on. xt = Σεi + x0 . zt = Σν i + z0 . 10 . Sargan (1964) linked static-equilibrium economic theory to dynamic empirical models by embedding (18) in an autoregressive-distributed lag model: yt = b0 + b1 yt−1 + b2 xt + b3 xt−1 + εt . Section 5 replicates a simulation experiment that Yule undertook. then yt = β 0 + ut . Moreover. I(1). and α0 +α2 β 0 = b0 . Thus. However. xt is aggregate income. Yule’s message acted as a signiﬁcant discouragement to time-series work in economics. In this case. and ut would be stationary unless there were other non-stationary variables than income causing consumption. his three types of series are called integrated of orders zero. ensuring nonstationarity resulted in empirical economists usually making careful adjustments for factors such as population growth and changes in the price level. β 1 = (b2 +b3 )/(1−b1 ). Indeed. Diﬀerencing an I(1) series delivers an I(0). Hendry (1980) constructed a nonsense regression by using cumulative rainfall to provide a better explanation of price inﬂation than did the money stock in the UK. Thus. Models such as (20) explain growth rates in yt by the growth in xt and the 7 Recall lower case letters are in logaritmic form. Consequently. with ut being non-stationary as the extreme case. and two respectively (I(0). one. As economic variables have trended over time since the Industrial Revolution.7 If aggregate income is linearly related to aggregate consumption in a causal way. unrelated to xt .e. A technical analysis of the sources and symptoms of the nonsense-regressions problem was ﬁnally presented by Phillips (1986). and the latter is a random walk. Assume that yt is aggregate consumption.e. Yule’s problem in case B can be seen clearly: if β 1 were zero in (18). Nevertheless. as before.Yule found that rxy was almost normally distributed in case A.. if β = 0 and ut is stationary in the regression equation: yt = β 0 + β 1 xt + ut (18) then yt and xt must contain the same stochastic trend. i. since otherwise ut could not be stationary. ut contains Σν i so is non-stationary and. it seemed as if inference could go badly wrong once the data were non-stationary. i.. He was startled to discover that rxy had a U-shaped distribution in C. say. therefore. (19) The dynamic model (19) can also be formulated in the so-called equilibrium-correction form by subtracting yt−1 from both sides and subtracting and adding b2 xt−1 to the right-hand side of (19): ∆yt = α0 + α1 ∆xt − α2 (yt−1 − β 1 xt−1 − β 0 ) + εt (20) where α1 = b2 . but became nearly uniformly distributed (except at the end points) in B. ut being autocorrelated in (18). inconsistent with the stationarity assumption of the regression model. yt is non-stationary. then yt would ‘inherit’ the non-stationarity from xt . and thereby implicitly assumed constant proportional relations between non-stationary variables. is what induces the non-standard distributions of rxy . Assume now that. Today. all coeﬃcients in (20) can be derived from (19). but gradually its impact faded. they often worked with the logarithms of data (to ensure positive outcomes and models with constant elasticities). one cannot conduct standard tests of the hypothesis that β 1 = 0 in such a setting. but it is not caused by xt and instead is determined by another non-stationary variable. α2 = (1−b1 ). For example. β 1 = 0 in (18) is the correct hypothesis. so the correct null hypothesis (of no relation between x and y) was virtually certain to be rejected in favor of a near-perfect positive or negative link. and hence what one would like to accept in statistical tests.

But from (20). therefore. and argued that terms like ut−1 would often be stationary even when the individual series were not. but also as a result of previous period’s consumption not being in equilibrium (i. Srba and Yeo (1978) introduced a class of models based on (20) which they called ‘error-correction’ mechanisms (denoted ECMs). but drifts as a non-stationary variable.. When α2 = 0. no longer trending. hence we can derive that α0 = −(α1 − β 1 )gx . Notice that when εt . The latter case can be seen from (19). Hendry. we can derive the properties of the equilibrium error ut = (y − β 1 x − β 0 )t : (y − β 1 x − β 0 )t = α0 + (α1 − β 1 )∆xt + (1 − α2 )(y − β 1 x − β 0 )t−1 + εt or: ut = ρut−1 + α0 + (α1 − β 1 )∆xt + εt .past disequilibrium between the levels. Notice also that by subtracting β 1 ∆xt from both sides of (20) and collecting terms. unlike the ‘nonsense regressions’ case. when ut−1 ∼ I(0). More formally. yt−1 = β 0 + β 1 xt−1 ). as occurs in (21). To understand the status of equations like (20). Standard t-statistics will be ‘sensibly’ distributed (assuming that εt is IID). 9 The resulting distribution is not actually a t-statistic as proposed by Student (1908): Section 6 shows that it depends in part on the Dickey–Fuller distribution. Granger (1981) introduced the concept of cointegration where a genuine relation exists. However. cannot be explained by a non-stationary variable. I(1) ‘nonsense-regressions’ problems will disappear in (20) because ∆yt and ∆xt are I(0) and. the ‘equilibrium error’ ut = (y − β 1 x − β 0 )t is a stationary autoregressive process. is stationary or not. there is no adjustment. t is well behaved. the higher the value of ρ (equivalently the smaller the value of α2 ). If α2 = 0. Conversely. the slower is the adjustment back to equilibrium. Hendry and Anderson (1977) noted that ‘there are ways to achieve stationarity other than blanket diﬀerencing’. and by integrating we get yt = α1 xt + εt . Based on equations like (20). and the longer it takes for an equilibrium error to disappear. For example. ∆yt and ∆xt are I(0). The magnitude of the past disequilibrium is measured by (yt−1 − β 1 xt−1 − β 0 ) and the speed of adjustment towards this steady-state by α2 . it has to be corrected downwards. gy = α0 + α1 gx . ∆yt . despite the non-stationary nature of the original data. the equilibrium error is an autocorrelated process. where ρ = (1 − α2 ). Think of a situation where consumption changes (∆yt ) as a result of a change in income (∆xt ). or if it was too low. and yt does not return to any equilibrium value. if previous consumption was too high. then α2 measures the speed of adjustment with which ∆yt adjusts (corrects) towards each new equilibrium position. ut−1 . it has to be corrected upwards. Davidson. then ∆yt = α0 +α1 ∆xt +εt .9 This is because a stationary variable. To summarize: when α2 = 0 (so ρ = 1). Further evidence that many economic time series were better construed as non-stationary than stationary was presented by Nelson and Plosser (1982). Engle and Granger (1987) proved that ECMs and cointegration were actually two names for change in the equilibrium yt = β 1 xt − β 0 is ∆yt = β 1 ∆xt . or • α2 = 0 and (yt−1 − β 1 xt−1 − β 0 ) ∼ I(1). Closing this circle.e. irrespective of whether the past equilibrium error.8 Thus. there are two possibilities: • α2 = 0 and (yt−1 − β 1 xt−1 − β 0 ) ∼ I(0). who tested for the presence of unit roots and could not reject that hypothesis. and α2 0 if ut−1 ∼ I(1). so these variables must have steady-state growth rates gy and gx related by gy = β 1 gx . 8 The (21) 11 . thereby introducing the obverse of nonsense regressions.

(29) 10 Some references to the vast literature on the theory and practice of testing for both unit roots and cointegration include: Banerjee and Hendry (1992). 1989). Johansen (1988.the same thing: cointegration entails a feedback involving the lagged levels of the variables. where the data are cumulated zero. and conducting a t-test for no relationship. Banerjee. 1992a. Hendry (1995). namely: ∆yt = t where t ∼ IN 0. 12 .10 5 Nonsense regression illustration Using modern software. the t-statistic satisﬁes: P tβ 1 =0 ≥ 2. and a lagged feedback entails cointegration. s. and Stock (1987). Phillips and Perron (1988). Also: E [ t ν s ] = 0 ∀t. 1992). 1981). say) is obtained by dividing the estimated coeﬃcient by its standard error: tβ1 =0 = where: β1 = and SE β 1 = σu . once and twice. σ2 ν and setting y0 = 0. The process used to generate the I(1) data mimics that used by Yule. σ 2 (22) (23) ∆xt = ν t where ν t ∼ IN 0. it is easy to demonstrate nonsense regressions between unrelated unit-root processes. Johansen and Juselius (1990. 1987b. A t-test of H0 : β 1 = 0 (as calculated by a standard regression package. 1995b). x0 = 0. Park and Phillips (1988. 1987a. (27) When ut correctly describes an IID process. Dickey and Fuller (1979. Hall and Heyde (1980). Galbraith and Hendry (1993). Phillips (1986. Chan and Wei (1988). 1991. 1988). (xt − x)2 (28) (xt − x)2 −1 β1 SE β 1 (26) (xt − x)(yt − y). before regressing either series on the other.0 | H0 0. The economic equation of interest is postulated to be: yt = β 0 + β 1 xt + ut where β 1 is believed to be the derivative of yt with respect to xt : ∂yt = β1. 1992b. Yule’s three cases correspond to generating uncorrelated bivariate time series.05. ∂xt (24) (25) Equations like (25) estimated by OLS wrongly assume {ut } to be an IID process independent of xt . Dolado.

8 to deﬁne a 5% rejection frequency under the null because: P tβ 1 =0 ≥ 14. Thus: (yt − yt )2 ≤ (yt − y)2 .This.e. The ﬁrst panel (a) shows the distribution of the t-test on the coeﬃcient of xt in a regression of yt on xt when both variables are white noise and unrelated. inducing many big ‘t-values’. (30) where yt = β 0 + β 1 xt . from the Monte Carlo experiment. will produce smaller residuals. This is because the estimated value β 1 is usually diﬀerent from zero (sometimes u widely so) and. so E[β 1 ] = 0. yt−1 and xt−1 when the data are generated as unrelated stationary ﬁrst-order autoregressive processes. as is evident from the graphs in Figures 1 and 3. we should use 15. it is the deviation from the previous value that measures the stochastic change in xt . This is so because x (instead of xt−1 ) is a very poor ‘reference line’ when xt is trending. From (28). if ut is I(1). In fact.. whereas when xt is non-stationary. so serious over-rejection occurs using (29). the sum of squares (xt − x)2 is not an appropriate measure of the variance in xt when xt is non-stationary. SE[β 1 ] consists of two components. however. hence. The ﬁrst and third panels are close to the actual distribution of Student’s t. the dispersion of β 1 around zero is also large. It is now easy to understand the outcome of the simulation study: because the correct standard SE[β 1 ]. at T = 100. the deviation from the mean is a good measure of how much xt has changed. σu . (31) so both (30) and (31) work in the same direction of producing a serious downward bias in the estimated value of SE[β 1 ]. Figure 4 reports the frequency distributions of the t-tests from a simulation study by PcNaive (see Doornik and Hendry. The ﬁnal panel (d) shows the t-test on the equilibrium-correction coeﬃcient α2 in (20) for data generated by a cointegrated process (so α2 = 0 and β is known). Why does such a large distortion occur? We will take a closer look at each of the components in (26) and expand their formulae to see what happens when ut is non-stationary. Therefore: (xt − x)2 (xt − xt−1 )2 . is not the case if ut is autocorrelated: in particular. the former is as expected from statistical theory. the estimated residual variance σ 2 will in general be lower than u σ 2 = (yt − y)2 /T. which is the approximate 95% conventional conﬁdence interval. The intuition is that although β 1 is an unbiased estimator of β 1 . i. with a distributional spread so wide that 13 . and our artiﬁcial example 2. More importantly. we would need a critical value of 14. The third panel (c. 000 drawings for T = 50. The shaded boxes are for ±2. When β 1 = 0. This is numerically very close to the correct distribution of a t-variable. 1998). σ β1 and negative values both occur. whereas the latter shows that outcome is approximately correct in dynamic models once the dynamics have been included in the equation speciﬁcation. (xt − x)2 .05. big positive error is extremely large. The second panel (denoted b.8 | H0 0. and the sum of squares. When the data are stationary. using M = 10. in the top row) shows the equivalent distribution for the nonsense regression based on (22)–(26). the residual standard error. it has a very large variance. but the calculated SE[β 1 ] dramatically under-estimates the true value. The second panel shows an outcome that is wildly diﬀerent from t. left in the lower row) is for the distribution of the t-test on the coeﬃcient of xt in a regression of yt on xt . Instead of the conventional critical value of 2.

1 −4 −2 0 2 4 0 2 4 6 8 Figure 4: Frequency distributions of nonsense-regression t-tests most of the probability lies outside the usual region of ±2.4 . t-values. Thus. Even in the best case.. when β 1 = 0.4 −2 0 2 4 . Since the computer will calculate the coeﬃcients independently of whether the variables are stationary or not (and without issuing a warning when they are not).075 nonsense regression . While the last distribution is not centered on zero – because the true relation is indeed non-null – it is included to show that the range of the distribution is roughly correct.025 −4 . neither dynamics nor unit-root non-stationarity induce serious distortions: the nonsense-regressions problem is due to incorrect model speciﬁcation. so the excess rejection is due to the wrong standard error being used in the denominator (which as shown above.3 cointegrated relation . The ﬁrst point means that one will too frequently reject the null hypothesis (β 1 = 0) when it is true. nevertheless tβ 1 =0 diverges to inﬁnity as T increases.. and R2 based on OLS. 14 .1 .2 . Indeed.3 2 unrelated stationary variables .5 −20 −10 0 10 20 30 .1 . the correct distribution results for tβ 1 =0 . is badly downwards biased by the untreated residual autocorrelation).2 .e. when yt and xt are causally related. (ii) R2 cannot be interpreted as a measure of goodness-of-ﬁt. when it is true. panels (c) and (d) show that. i. Almost all available software packages contain a regression routine that calculates coeﬃcient estimates. 1995. when yt−1 and xt−1 are added as regressors in case (b). standard t-tests will be biased with too frequent rejections of a null hypothesis such as β 1 = 1. by themselves.05 . chapter 3).4 .3 2 unrelated white−noise variables .2 . so that conventionallycalculated critical values are incorrect (see Hendry. it is important to be aware of the following implications for regressions with trending variables: (i) Although E[β 1 ] = 0. delivering a panel very similar to (c).

reformulating the model to have white-noise errors is a good step towards solving the problem. Consider estimating β in the autoregressive model: yt = βyt−1 + t where t ∼ IN 0. the form of the limiting distribution of the DF test is also altered by the presence of a constant or a trend in either the DGP or the model. even after such a scaling. We will now discuss how to test for the presence of unit roots in the data. is easily computed. As demonstrated in Section 2.Hence statistical inference from regression models with trending variables is unreliable based on standard OLS output. wrong choices of what deterministic terms to include – or which table is applicable – can seriously distort inference. If autocorrelation is found. using a sample of T 2 size T . So even though unit roots impart an important nonstationarity to the data. All this points to the crucial importance of always checking the residuals of the empirical model for (unmodeled) residual autocorrelation. but mistakes are more easily made when the data are non-stationary. Thus. the role and the interpretation of the constant term and the trend in the model changes as we move from the stationary case to the non-stationary unit-root case. and using them can lead to over-rejection of the null of a unit root when it is true. the data second moments (like t=1 yt−1 ) grow at order T 2 . 400. because many of the conventional statistical distributions. often called the Dickey–Fuller test after Dickey and Fuller (1979). The second point will be further discussed and illustrated in connection with the empirical analysis. This means that diﬀerent critical values are required in each case. It is also the case for stationary data that incorrectly omitting (say) an intercept can be disastrous. it is often a good idea to act as if there are unit roots to obtain robust statistical inference.. the form of the limiting distribution is diﬀerent from that holding under stationarity. the F. 6 Testing for unit roots We have demonstrated that stochastic trends in the data are important for statistical inference. 100. scaling by T is required (rather than T for I(0) data).95). and the χ2 distributions become approximately valid once the model is re-speciﬁed to have a white-noise error. although all the required tables of the correct critical values are available. no determinist trend in the levels). For comparison. Unfortunately. but with a root close to unity (say. the Augmented Dickey-Fuller test is deﬁned in the next section. ρ > 0. Even though a variable is stationary. Again. conventional critical values are incorrect. Moreover. Rather. 15 . the distinction between a unit-root process and a near unit-root process need not be crucial for practical modeling. Worse still. We will now consider unit-root testing in a univariate setting. An example of this is given by the empirical illustration in Section 8. so the distribution of β − β ‘collapses’ very quickly. but does not have a standard t-distribution. and transforming the variables to be I(0) will complete the solution. making it hard to discriminate the null of a unit root from alternatives close to unity. Because V [yt ] = σ 2 t. such as Student’s t. The more general test.e. the Dickey–Fuller (DF ) test has a skewed distribution with a long left tail. Consequently. The four panels for the estimated distribution in Figure 5 have been standardized to the same x-axis for visual comparison – and the convergence is dramatic. we can illustrate this by simulation. However. The ‘t-statistic’ for testing H0 : β = 1. σ2 (32) under the null of β = 1 and y0 = 0 (i. to obtain a limiting distribution for β − β which neither diverges nor degenerates √ to a constant. where second moments grow at order T . estimating (32) at T = 25. the corresponding graphs for β = 0.5 are shown in Figure 6. and 1000. then the model should be re-speciﬁed to account for this feature.

3 T=25 2 10 T=100 5 1 −.25 . If xt ∼ I(1). assuming that xt can be treated as weakly exogenous for the parameters of the conditional model (see e. there is no guarantee that yt − β 1 xt − β 0 is I(0) for any value of β. 1983). as the discussion in Section 4 demonstrated..25 Figure 5: Frequency distributions of unit-root estimators However. provided the correct model is used (see discussion in the next section). These procedures will be described in Part II.75 1 1.25 0 .11 As discussed in Section 5. 1981).75 1 1. as in xt − xt−1 .75 1 1.25 50 40 30 20 10 0 . But unlike diﬀerencing..5 .25 . apart from that corresponding to a test for a unit root. There are many possible tests for cointegration: the most general of them is the multivariate test based on the vector autoregressive representation (VAR) discussed in Johansen (1988). PcGive) automatically provides appropriate critical values for unit-root tests in almost all relevant cases. for example.g. which induces cointegration when yt − β 1 xt − β 0 ∼ I(0).75 1 1.5 . then by deﬁnition ∆xt ∼ I(0).5 . and most time-series econometric software (e. Engle. 11 Regression 16 . 7 Testing for cointegration When data are non-stationary purely due to unit roots.25 −.g. they can be brought back to stationarity by linear transformations.25 . by diﬀerencing. Most tests still have conventional distributions. The required critical values have been tabulated using Monte Carlo simulations by Dickey and Fuller (1979. always including a constant and a trend in the estimated model ensures that the test will have the correct rejection frequency under the null for most economic time series.25 .25 T=400 100 T=1000 50 −. cointegrated).. Here we only consider tests based on the static and the dynamic regression model. the condition that there exists a genuine causal link between I(1) series yt and xt is that the residual methods can be applied to model I(1) variables which are in fact linked (i.25 −.e. Hendry and Richard.25 0 . An alternative is to try a linear transformation like yt − β 1 xt − β 0 .5 .25 0 .

5 −.25 0 .75 1 Figure 6: Frequency distributions of stationary autoregression estimators ut ∼ I(0). or equivalently. that ρ < 1 in (2). 17 (35) .75 1 −. H0 : 1 − ρ = 0 in: ut = ρut−1 + εt or: ∆ut = (1 − ρ)ut−1 + εt .5 . implies using a non-standard distribution.25 .5 T=400 10 5 5 T=1000 2. Therefore. the Engle–Granger test procedure is based on testing that the residuals ut from the static regression model (18) are stationary.25 .5 .25 0 .25 .e. otherwise a ‘nonsense regression’ has been estimated.25 0 . for a discussion of this drawback) is that the autoregressive model (33) for ut is the equivalent of imposing a common dynamic factor on the static regression model: (1 − ρL)yt = β 0 (1 − ρ) + β 1 (1 − ρL)xt + εt .25 0 . then the null hypothesis of the DF test is H0 : ρ = 1. using the DF test. the test of the null of a unit coeﬃcient.5 . A drawback of the DF -type test procedure (see Campos.5 1 . Ericsson and Hendry. The test is based on the assumption that εt in (33) is white noise.5 .25 .5 −. 1996.5 15 −. As discussed in the previous section. then it has to be augmented by lagged diﬀerences of residuals: ∆ut = (1 − ρ)ut−1 + ψ 1 ∆ut−1 + · · · + ψ m ∆ut−m + εt . Let ut = yt − β 1 xt − β 0 where β is the OLS estimate of the long-run parameter vector β. and if the AR(1) model in (33) does not deliver white-noise errors.5 −.75 1 −. (34) (33) We call the test of H0 : 1−ρ = 0 in (34) the augmented Dickey–Fuller test (ADF ).5 4 T=25 3 2 1 T=100 −..5 −.75 1 7.2 1. i.

8 0 An empirical illustration Gasoline price A 0 −.25 0 1 1995 4 1990 3 2 1 −1 −.2 0 1995 Gasoline price B Density of A Density of B Correlogram of A Correlogram of B . its distribution is much closer to the Student t-distribution than the Dickey–Fuller. and correspondingly Banerjee et al. Although non-standard.t ) at diﬀerent locations over the period 1987 to 1998. The 18 . The test reported in the empirical application in the next section is based on this test. Empirical evidence has not produced much support for such common factors. we will apply the concepts and ideas discussed above to a data set consisting of two weekly gasoline prices (Pa. β 0 .75 1990 3 2 1 −1 1 −.e. Instead.. and its distribution is illustrated in Figure 4. so its critical values have been separately tabulated. panel d.5 .5 −.5 −.t and Pb.25 −. (36) where the parameters {α0 . α2 . In this case. their empirical densities and autocorrelograms In this section. Kremers.4 −. when not only yt adjusts to the previous equilibrium error as in (36). a multivariate test procedure is needed. (1993) ﬁnd the power of tα2 =0 can be high relative to the DF test. but also xt does). However. the test is potentially a poor way of detecting cointegration. the null rejection frequency of their test depends on the values of the ‘nuisance’ parameters α1 and σ 2 . Ericsson and Dolado (1992) contrast them with a direct test for H0 : α2 = 0 in: ∆yt = α0 + α1 ∆xt + α2 (yt−1 − β 1 xt−1 − β 0 ) + εt .5 −. α1 . when xt is not weakly exogenous (i.75 −. β 1 } are not constrained by the common-factor restriction in (35).6 −.8 −. the common-factor restriction in (35) should correspond to the properties of the data.5 0 5 10 0 5 10 Figure 7: Gasoline prices at two locations in (log) levels. Unfortunately. rendering such tests non-optimal. so Kiviet and Phillips (1992) developed a test which is invariant to these ε values.For the DF test to have high power to reject H0 : 1 − ρ = 0 when it is false.

2 .1 1 0 .05 0 . suggesting non-stationarity. 1 − a2 1 − a2 19 (40) t (39) . and: β0 = a0 a1 + a3 and β 1 = . we will assume that E[∆yt ] = E[∆xt ] = 0. The bimodal frequency distribution of the price levels is also typical of non-stationary data.e. i. suggesting stationarity (the latter are shown with lines at ±2SE to clarify the insigniﬁcance of the autocorrelations at longer lags).05 .2 1 −. perhaps with a couple of outliers. (38) Consistent with the empirical data.1 . and the lack of such autocorrelations for the diﬀerenced prices. in a typically stationary manner.1 1990 20 1995 0 −.1 Gasoline price B 1995 Density of A Density of B 10 10 −. Without changing the basic properties of the model. The price levels exhibit ‘wandering’ behavior. whereas the frequency distribution of the diﬀerences is much closer to normality.. and in (log) diﬀerences in Figure 8. and then consider the linear dynamic model: yt = a0 + a1 xt + a2 yt−1 + a3 xt−1 + t .1 .15 Correlogram of A Correlogram of B 0 0 0 5 10 0 5 10 Figure 8: Gasoline prices at two locations in diﬀerences.data in levels are graphed in Figure 7 on a log scale.1 1990 20 . their empirical density distribution and autocorrelogram We ﬁrst report the estimates from the static regression model: yt = β 0 + β 1 xt + ut . We also notice the large autocorrelations of the price levels at long lags.1 −. Gasoline price A . though not very strongly.1 0 −. whereas the diﬀerenced series seem to ﬂuctuate randomly around a ﬁxed mean of zero. we can then rewrite (38) in the equilibrium-correction form: ∆yt = a1 ∆xt + (a2 − 1) (y − β 0 − β 1 x)t−1 + where a2 = 1. there are no linear trends in the data.

Although the DW test statistic is small and suggests nonstationarity. The residuals show substantial temporal correlation. In (39). since yt − β 0 − β 1 xt = ut ∼ I(0) implies a well-behaved equilibrium. it tests for skewness and excess kurtosis of the residuals. In 20 .050. It is based on the third and the fourth moments around the mean.4) (42) R = 0. (41) whereas if (a2 − 1) = 0. distributed as χ2 (2). To evaluate the forecasting performance of the model. the large sample size improves the precision of the test and. ∆xt ) are I(0) when their corresponding levels are I(1).6 [0. F(7. though declining. and the ‘t-statistics’ based on (26) are given in parentheses. T ). the normality and homoscedasticity of the residuals are also rejected. As demonstrated in (21). in (b) the residuals ut . the model embodies the lagged equilibrium error (y − β 0 − β 1 x)t−1 . Normality denotes the Doornik and Hansen (1994) test of residual normality. This is further conﬁrmed by the correlogram.In formulation (39). This type of ‘balancing’ occurs naturally in regression analysis when the model formulation permits it: we will demonstrate empirically in (43) that one does not need to actually write the model as in (39) to obtain the beneﬁts. i.predictions.t = 0. F(7. we have calculated the one-step ahead forecasts and their (calculated) 95% conﬁdence intervals over the last two years. The test of autocorrelated errors of order 1-7 is very large and the null of no autocorrelation is clearly rejected.00]∗∗ 22. hence. so with t ∼ I(0).2) (67. which shows periods of consistent over. The estimates of the static regression model over 1987(24)–1998(29) are: pa.89. The ARCH (m) (see Engle.2 [0. The outcome is illustrated in Figure 10a. Furthermore. Furthermore.18 where DW is the Durbin–Watson test statistic for ﬁrst-order autocorrelation. no equilibrium exists. 569) = ARCH (7). (∆yt . in (c) their correlogram and in (d) the residual histogram compared with the normal distribution. allows us to reject the hypothesis.and under. i. σ u = 0. So the standard assumptions underlying the static regression model are clearly violated. 1982) is a test of autoregressive residual heteroscedasticity of order m distributed as F(m. T − m). What matters is whether the residuals are uncorrelated or not.00]∗∗ 2 The AR(1–m) is a test of residual autocorrelation of order m distributed as F(m.018 + 1. 562) = Normality .21∗∗). autocorrelations. consistent with a highly autocorrelated process. DW = 0. but are leptokurtic to some extent. the residuals seem to be symmetrically distributed around the mean. The link of cointegration to the existence of a long-run solution is manifest here. there is no adjustment back to equilibrium and the equilibrium error is a non-stationary process. whereas when ut ∼ I(1). This can explain why the DF test rejected the unit-root hypothesis above: though ρ is close to unity. which captures departures from the long-run equilibrium as given by the static model. exhibiting large. the equilibrium error will be a stationary process if (a2 − 1) = 0 with a zero mean: E [yt − β 0 − β 1 xt ] = 0. panel (a). a test of H0 : ut = εt against H1 : ut = ρ1 ut−1 + · · · + ρm ut−m + εt . χ2 (2) = 524.9 [0. the following mis-speciﬁcation tests were calculated: AR(1–7).00]∗∗ 213.e.t + ut (2. Thus.e. the DF test of ut in (42) supports stationarity (DF = −8.01 pb.. In Figure 9. we have graphed the actual and ﬁtted values from the static model (42). the equation is ‘balanced’ if and only if (y − β 0 − β 1 x)t is I(0) as well.

at the end of 1997 until the beginning of 1998.91 pb.5 1995 1990 1995 Residual correlogram .1) (12. so values > 1 reject).t − 1.050 to 0. predictions were consistently below actual prices.12 Finally.34 pb.001 + 1.t = 0. which is ﬂagrantly wrong here.4) (43) R = 0.t−1 + 0.9) (14. We also report the 1-step residuals with ±2SE. and thereafter consistently above. so the precision has increased 12 The software assumes the residuals are homoscedastic white-noise when computing conﬁdence intervals. most empirical economists would consider this econometric outcome as quite unsatisfactory.0 −. their correlogram and histogram.018.t−2 + εt (0. etc. It appears that even if the recursive estimates of β 0 and β 1 are quite stable. Figure 11 shows these recursive graphs. We will now demonstrate how much the model can be improved by accounting for the left-out dynamics.99. standard errors.12 pb. residuals.03 Compared with the static regression model.. we have recursively calculated the estimated β 0 and β 1 coeﬃcients in (42) from 1993 onwards. The estimate of the dynamic regression model (38) with two lags is: pa.25 −. Altogether.25 0 5 10 −4 −2 0 2 4 Figure 9: Actual and ﬁtted values. and the sequence of constancy tests based on Chow (1964) statistics (scaled by their 1% critical values.t−2 + 0.75 . σ ε = 0.t−1 − 0.75 Actual and fitted values Actual for A Fitted 4 Residuals 2 0 −2 1990 1 .5) (23. DW = 2.1) .5 −.3) (36. we notice that the DW statistic is now close to 2. 2 21 .33 pa. at the end of the period they are not within the conﬁdence band at the beginning of the recursive sample (and remember the previous footnote).4 Residual density N(0.2 .018.8) (6. from the static model particular. and that the residual standard error has decreased from 0.46 pa.

F(7.25∗∗ (0.t = 0. residuals. 563) = 1.or overprediction in the dynamic model. Figure 12 show the graphs of actual and ﬁtted. This is also evident from the histogram exhibiting quite long tails. However. residual correlogram.t approximately 2. prevailing approximately till the end of 1992.8 [0.25 95% confidence bands −. probably due to the Gulf War.2 .t ecmt−1 1990 1995 1990 1995 Figure 10: Static and dynamic model forecasts. 556) = 10. Fitted and actual values are now so close that it is no longer possible to visually distinguish between them.t tur = 8.1 0 −. The rejection of homoscedastic residual variance seems to be explained by a larger variance in the ﬁrst part of the sample. but there is still evidence of the residuals being heteroscedastic and non-normal. with the increased precision (the smaller residual standard error). χ2 (2) = 79. As for the static regression model.8) (44) 22 . It appears that there is no systematic under.5 −.1 ∆pa. F(7. the equilibrium error. and the residual histogram compared to the normal distribution.19] ARCH (7).99 pb. Also the prediction intervals are much narrower.3) (22. It is now possible to derive the static long-run solution as given by (39) and (40): pa.6 [0. resulting in the rejection of normality.1 Actual Static model fitted Static model forecasts 1997 1998 −.75 Actual Dynamic model fitted Dynamic model forecasts 1996 1997 1998 Dynamic model equilibrium error . the one-step ahead prediction errors with their 95% conﬁdence intervals have been graphed in Figure 10b. we can now recognize several ‘outliers’ in the data.25 −.5 −.75 1996 . Residuals exhibit no temporal correlation.008 + 0.1 0 −.5 times.00]∗∗ Normality . and its relation to ∆pa. The mis-speciﬁcation tests are: AR(1–7).2 . reﬂecting the large increase in precision as a result of the smaller residual standard error in this model.44 [0.00]∗∗ We note that residual autocorrelation is no longer a problem.95% confidence bands −.

we ﬁnd that the autocorrelation coeﬃcient (1 − α2 ) in (21) corresponds to 0. revealing an insigniﬁcant intercept. In the static model. though one is formulated in non-stationary and the other in stationary variables. consistent with (21).34 ∆pb.3) (6. consistent with the downward bias of SE[β] in the static regression model.75 .5) (24. Note that ut has a zero mean.05 Notice that the corresponding estimated coeﬃcients.5 .91 ∆pb.1 1−step residuals with± 2SE 1990 1995 1990 1995 Figure 11: Recursively-calculated coeﬃcients of β 0 and β 1 in (41) with 1-step residuals and Chow tests Note that the coeﬃcient estimates of β 0 and β 1 are almost identical to the static regression model. However. pa. the correctly-calculated standard errors of estimates produce much smaller t-values.t−1 − 0. even though the test for two common factors in (43) rejects (one common factor is accepted). demonstrating that the two models are statistically equivalent.3) (45) R = 0.2 Recursively−computed constancy test . the values of the DF test on the static residuals. The graph of the equilibrium error ut (shown in Figure 10c) is essentially the log of the relative price. consistent with (41).13 ecm t−1 + εt (12.1 1. DW = 2.. and that it is strongly autocorrelated.t − 0.86.9 Recursively−computed slope with ±2SE.t − pb. Finally.8 1990 1 1995 . we report the estimates of the model reformulated in equilibrium-correction form (39).1 . whereas in the dynamic model. which is a fairly high autocorrelation.46 ∆pa.05 −.1 1 Recursively−computed intercept with ± 2SE. static model . static model 1990 1995 .25 0 −. the residual standard error. The coeﬃcient 23 2 .69. From the coeﬃcient estimate of the ecm t−1 below in (45).t−1 + 0.05 0 −. Nevertheless.t = 0. suppressing the constant of zero due to the equilibrium-correction term being mean adjusted: ∆pa. and DW are identical with the estimates in (43). we have explained it by short-run adjustment to current and lagged changes in the two gasoline prices. and the unit-root test in the dynamic model (tur in (44)) are closely similar here.t . illustrating the fact that the OLS estimator was unbiased.018.1 .5) (8. σ ε = 0. this autocorrelation was left in the residuals.

residuals. Finally. demonstrating the hazards of interpreting R2 as a measure of goodness of ﬁt. with 13% of any disequilibrium in the relative prices being removed each week. so that aspect of weak exogeneity is not rejected. In contrast. Hence. R2 is lower in (45) than in (42) and (43)..2 −. which explains the high R2 often obtained in regressions with non-stationary variables.5 −.t in a bivariate system.5 0 .75 Actual Fitted Residual 2. 24 . Such non-constancies reveal remaining non-stationarities in the two gasoline prices.13 on ecm t−1 suggests moderate adjustment. 13 The equilibrium-correction error ecm t−1 does not inﬂuence ∆pb.0 −. we essentially compare how well our model can predict yt compared to a straight line.t to ecm t−1 . When we calculate R2 = Σ(yt − yt )2 /Σ(yt − y)2 . although a few of the recursive estimates fall outside the conﬁdence bands (especially around the Gulf War).. The estimated coeﬃcients are reasonably stable over time.4 Density N(0. but resolving this issue would necessitate another paper. any other trending variable can do better than a straight line.25 −.. However.. we report the recursively-calculated coeﬃcients of the model parameters in (45) in Figure 13. the sum of squares Σ(yt − y)2 is not an appropriate measure of the variation of a trending variable.5 0 −2. R2 = Σ(∆yt − ∆yt )2 /Σ(∆yt − ∆y)2 from (45) measures the improvement in model ﬁt compared to a random walk as a reasonable measure of how good our model is (although that R2 would also change if the dependent variable became ecm t in another statistically-equivalent version of the same model).13 Figure 10d shows the relation of ∆pa. their correlogram and histogram for the dynamic model of −0. as already discussed in Section 5.5 1990 1 1995 1990 1995 Correlogram . When yt is trending.5 .1) 0 5 10 −4 −2 0 2 4 Figure 12: Actual and ﬁtted values.

and some theoretical models entail unit roots. Thus. where the latter comprised linear combinations of the variables that did not have unit roots. For modeling purposes.t−1 . and inference when the stationarity assumption is incorrect can induce serious mistakes. Unit-root non-stationarity seems widespread in economic time series. empirical evidence is strongly against the validity of that assumption.25 1 .8 1.4 1990 . We investigated the comparative properties of stationary and non-stationary processes. Next. reviewed the historical development of modeling non-stationarity and presented a re-run of a famous Monte Carlo simulation study of the dangers of ignoring non-stationarity in static regression analysis. but could be transformed back to stationarity by diﬀerencing and cointegration transformations.6 . focusing on autoregressive processes with unit roots.3 1990 1995 Figure 13: Recursively-calculated parameter estimates of the ECM model 9 Conclusion Although ‘classical’ econometric theory generally assumed stationary data. we described how to test for unit roots in scalar autoregressions. an extensive empirical illustration using two gasoline prices implemented the tools described in the preceding analysis. a unit-root process may also be considered as a statistical approximation when serial 25 .25 −.2 1995 .t−1 0 ecmt−1 −. Cointegrated relations and diﬀerenced data both help model unit roots.25 1990 1995 ∆pb.∆pa. We showed that these processes were non-stationary. then extended the approach to tests for cointegration.25 0 −.1 −. and can be related in equilibrium-correction equations. we believe it is sensible empirical practice to assume unit roots in (log) levels until that is rejected by well-based evidence. Nevertheless. particularly constant means and variances across time periods. stationarity is an important basis for empirical modeling.t .5 . we considered recent developments in modeling non-stationary data.5 −.75 1990 1995 −. Finally. Links between variables will then ‘spread’ such non-stationarities throughout the economy. as we illustrated.75 ∆pb. To develop a more relevant basis.

187–220.S. 1057–1072. C. Dolado. H. S. there is excellent software available for implementing the methods discussed in Johansen (1995a).. F. and Hendry. Z. and Hendry. Important new insights result. This emphasizes the advantages of accounting for the dynamic properties of the data in equilibrium-correction equations. which not only results in improved precision from lower residual variances.. Galbraith. F.. G. Chan. Unfortunately. 26 . and Fuller.. D. A. W. J.A. 74. D. and recent developments have focused on this approach. 532–553. M. Dickey. F. 1995) and PcFiml (see Doornik and Hendry. (1981).. Reprinted in Hendry. F. Co-integration. 59. Econometric modelling of the aggregate time-series relationship between consumers’ expenditure and income in the United Kingdom. other sources of non-stationarity may remain.. N. F. N. Ericsson. multivariate analysis is natural. Banerjee. A. Annals of Statistics. Fortunately. Hendry. Econometrics: Alchemy or Science? Oxford: Blackwell Publishers. Demery.. Likelihood ratio statistics for autoregressive time series with a unit root. and Fuller. preferably homoscedastic. and Hendry. We reiterate the importance of having white-noise residuals. C. Estimating the UK demand for money function: A test of two approaches. Chow.. D. Distribution of the estimators for autoregressive time series with a unit root. and Starr. Economic Journal. D. 367–401. (1988). W. such as changes in parameters (particularly shifts in the means of equilibrium errors and growth rates) or data distributions. J. (1996).. 1997). (1992). 54. Journal of the American Statistical Association. A comparison of alternative estimators for simultaneous equations. F. and we will address the application of such tools to the gasoline price data considered above. R. Journal of Econometrics. Since cointegration inherently links several variables.. and Wei. Hendry. A. 49. A.. J. J. Limiting distributions of least squares estimates of unstable autoregressive processes. References Attﬁeld. Baba. D. Error Correction and the Econometric Analysis of Non-Stationary Data. W. Mimeo. 225–255. 70. Srba. Cointegration tests in the presence of structural breaks. (1979). Dickey. Review of Economic Studies. C. Part II of our attempt to explain cointegration analysis will address system methods. Campos. (1993). D. A. 661–692. to avoid mis-leading inferences. R. (1993). Econometrica. A. L. 32.. 1960–1988. and Duck. so careful empirical evaluation of ﬁtted equations remains essential. D. (1964).. (1995).. (1992). Testing integration and cointegration: An overview. The demand for M1 in the U. but delivers empirical estimates of adjustment parameters. H. J. Monte Carlo studies have demonstrated that treating near-unit roots as unit roots in situations where the unit-root hypothesis is only approximately correct makes statistical inference more reliable than otherwise. Davidson. Economics department. E.. Oxford: Oxford University Press. W. 16. J. but new modeling decisions also have to made in practice. Banerjee.. 427–431.. Y. F. F. and Yeo. including CATS in RATS (see Hansen and Juselius. Oxford Bulletin of Economics and Statistics. N. 88. 25–61. D. Econometrica. D.correlation is high. (1978). University of Bristol.

Econometrics: Alchemy or Science? Oxford: Blackwell Publishers. (1981). C. Some properties of time series data and their use in econometric model speciﬁcation. Econometrics: Alchemy or science?. Engle.) Testing Exogeneity. Exogeneity. F. 251–276. A. (1995). C. F. J. J. (1998). Econometrics: Alchemy or Science? Oxford: Blackwell Publishers. and Anderson. and in Ericsson. and Interest Rates. (1988). pp. 88. Granger. F. (1995). Spurious regressions in econometrics. and Morgan. W. 50.. (1983). Reprinted in Hendry. (1980).Doornik. A practical test for univariate and multivariate normality. J. D. 3. J. and Richard. Vol. Hendry. 59. R. Friedman. (1982). and Granger. C. (1980). 47. 1993. A. S. Hendry. Econometrics: Alchemy or Science? Oxford: Blackwell Publishers. Hendry. P.. (1995). Journal of Econometrics. Reprinted in Hendry. Cambridge: Cambridge University Press. London: Timberlake Consultants Press. S.. P. R. Autoregressive conditional heteroscedasticity.. with estimates of the variance of United Kingdom inﬂations. F. A. Prices. F. F. G. Discussion paper.). D. (ed.. Journal of Econometrics. 1867–1975. Hall. IL. Hansen. Reprinted in Hendry. (1997). Engle. Frontiers in Quantitative Economics. W. F. 111–120. Economic Journal. 27 . and Schwartz. Econometrica. Econometrica. 55. and Juselius. (eds. D. (1977). Econometrica. Johansen. Nuﬃeld College. and Mizon. M. R. Granger. Econometrics: Alchemy or Science? Oxford: Blackwell Publishers. 549–563. Estima. D. University of Oxford. not a nuisance: A comment on a study of the demand for money by the Bank of England. estimation and testing. (1993). 361–383. W. Doornik. R. Testing dynamic speciﬁcation in small simultaneous systems: An application to a model of building society behaviour in the United Kingdom. Nuﬃeld College. M. (1993). D. S. Engle. 1551–1580. (1991). 121–130. and Newbold... Dynamic Econometrics. Monte Carlo simulation using PcNaive for Windows. Martingale Limit Theory and its Applications. Journal of Economic Dynamics and Control. Oxford: Oxford University Press. J. S. F. D. and Hansen. Serial correlation as a convenient simpliﬁcation. D. J.-F. CATS in RATS: Cointegration analysis of time series. (1974). (1993). (1978). C. 387–406. Monetary Trends in the United States and the United Kingdom: Their Relation to Income. 231–254. J. F. and Heyde. D. 12. Unpublished typescript. F.. 1994. J.. Modelling Dynamic Systems using PcFiml 9 for Windows. Hendry. H. Evanston. Amsterdam: North Holland Publishing Company. D. Cointegration and error correction: Representation. 2. and Hendry.. Chicago: University of Chicago Press. Hendry. J. H.. Estimation and hypothesis testing of cointegration vectors in Gaussian vector autoregressive models. Oxford: Oxford University Press. 987–1007. (1994). D. J. D. D. and Hendry. A. 16. F. E. Hendry. and Irons. (1987). The Foundations of Econometric Analysis. (1982). Reprinted in Hendry. C. M. London: Academic Press. 277–304. Econometrica. F. F. F. In Intriligator. Doornik.. F. Johansen. N. 51. Statistical analysis of cointegration vectors. G. D.. Discussion paper. K.. Economica.

311–340. Oxford Bulletin of Economics and Statistics. Maximum likelihood estimation and inference on cointegration – With application to the demand for money. Cambridge: Cambridge University Press. B. Johansen. S. D. G. (1993). 74. Phillips. S. B. Statistical Foundations of Econometric Modelling. 124–169 in Sargan J. Ericsson. S. 1. D. 55. (1986). J. N. 28 . J. J. B. Oxford Bulletin of Economics and Statistics. Phillips. Econometrica. S. Biometrika.. 277–301. and Juselius. 335–346.) (1984). F. Econometric Analysis for National Economic Planning. (1988). Phillips. 1021–1043. C. and Juselius. Johansen. Vol. Statistical inference in regressions with integrated processes. C. P. Johansen. and Phillips. In Hart. part 1. P. Mills. C. Econometric Theory. 56. M. (1987b). P. Determination of cointegration rank in the presence of a linear trend. C. Sargan. and Phillips. 33. 139–162. Econometric Theory. Kiviet. (1964). 25–63. A. A. R. A representation of vector autoregressive processes integrated of order 2. P.). P. Cambridge: Cambridge University Press. (1992a). 275–314 in Hendry D. K. (1988). and Plosser. S. Understanding spurious regressions in econometrics. 169–210. and as pp. B. F. Journal of Econometrics. (1989). I. C. M. E. P. 75. Journal of Monetary Economics. and Perron. Journal of Econometrics. Econometrics and Quantitative Economics. 5. Econometrica. Phillips. and Wallis K. (1988). pp... K. 54. Econometric Theory. Likelihood-based Inference in Cointegrated Vector Autoregressive Models. Y. and Whitaker. Wages and prices in the United Kingdom: A study in econometric methodology (with discussion). (1992). 211–244. The Formation of Econometrics: A Historical Perspective. (1992). 535–547. (1986). Park. (1995a). (1990). 188–202. 4. R. Johansen. C. (1992b). 8. part 2. Statistical inference in regressions with integrated processes. Park.. J. Oxford: Clarendon Press. Cambridge: Cambridge University Press. Likelihood based Inference on Cointegration in the Vector Autoregressive Model... Contributions to Econometrics.. J. Towards a uniﬁed asymptotic theory for autoregression. London: Butterworth Co. (1987a). 383–398. Oxford Bulletin of Economics and Statistics. Trends and random walks in macroeconomic time series: some evidence and implications. (1992). Reprinted as pp. Testing for a unit root in time series regression. P. Testing structural hypotheses in a multivariate cointegration analysis of the PPP and the UIP for UK. (1995b). The power of cointegration tests. and Phillips. J. 16 of Colston Papers. 349–367. Qin.. Phillips. Kremers.Johansen. Nelson. J. 95–131. K. The History of Econometric Ideas. J. Johansen. Oxford: Oxford University Press. C. Oxford: Oxford University Press. 468–497. Biometrika.. Exact similar tests for unit roots and cointegration. Regression theory for near-integrated time series. Oxford Bulletin of Economics and Statistics. 54. F. Time series regression with a unit root. P. 10. J. Spanos. 54. C.. 325–348. 53. B. Oxford: Basil Blackwell. B. D. (1988). B. G. S. S. (1990). Y. 52.. P. Morgan. and Dolado. (1982). C. (eds. D. Vol. (eds.

Yule. Journal of the Royal Statistical Society. G. H. J. (1926). Biometrika. Student (1908). 11–24. 55. 89. 1035–1056. A random diﬀerence series for use in the analysis of time series. Econometrica.Stock. (1987). 29 . Journal of the American Statistical Association. Asymptotic properties of least squares estimators of cointegrating vectors. Why do we sometimes get nonsense-correlations between time-series? A study in sampling and the nature of time series (with discussion). On the probable error of the mean. Working. (1934). 1–25. 29. H. 6. U. 1–64.