1

t;hapter 4
If you solve (A3.5) for A. it is possible to verify
,

TESTING FOR TRENOS ANO

UNIT ROOTS

The one·step ahead forecast of y is 1

y; =O-A.:)L!/1-1Y1-j
Fi

Since A 1 < 1, the summation is such that (1 - A 1 )IA.:'1 - 1 = l. Hence, the optima!
1 1

forecast of y 1 can be fonned as a geometrically weighted average of the past realiza-

tions of the series.
The Hodrick-Prescott Decomposition 1nspection of the auto orrelation function serves as a rough indicator of whether a
Another method of decomposing a series into a trend and stationary component has trend is present in a series. A slowly decaying ACF is indicative of a large charac-
been developed by Hodrick and Prescott (1984 ). Suppose you observe the values y 1 teristic root, true unit root process, or trend stationary process. Formal tests can
through Yr and want to decompose the series in to a trend { µ1 } and stationary com- help determine whether or not a system contains a trend and whether that trend is
ponen! y 1 - µ/' Consider the sum of squares deterministic or stochastic. However, the existing tests have little power to distin-
guish between near unit root and unit root processes. The aims of this chapter are
T T-1
to:
2
2
¡. (1IT)L,,(Y1 -µ1) +(A.!T)L[Cµ1+1 -µ1)-(µ1-µr-1)1 I. Develop and illustrate the Dickey-Fuller and augmented Dickey-Fuller tests for
J:1
;:.· t=I t=2 the presence of a unit root. These tests can also be used to help detect the pres-
¡...
i': .. ence of a deterministic trend. Phillips-Perron tests, which entail less stringent
: ¡: The problem is to select the {µ1 } sequence so as to minimize this sum of squares. restrictions on the error process, are illustrated.
¡1
1!;. In the minimization problem, A is an arbitrary constant reflecting the "cost" or
2. Consider tests for unit roots in the presence of structural change. Structural
penalty of incorporating fluctuations into the crend. In many applications, including
change can complicare the tests for trends; a policy regime change can result in
·. Hodrick and Prescott (1984) and Farmer (1993), A. is set equal to 1600. Increasing
the value of A acts to "smooth out" the crend. If A= O, the sum of squares is mini-
a structural break that makes an otherwise stationary series appear to be nonsta-
tionary.
mized when y 1 = µ1; the trend is egua! to y 1 itself. As A oo, the trend approaches a 3. Illustrate a general procedure to determine whether or nota series contains a unit
linear time trend. Intuitively, for large values of A, Hodrick-Prescott decomposition
forces the change in the trend (i.e., 6.µ 1+ 1 - tlµJ to be as small as possible. This oc- root. Unit root tests are sensitive to the presence of deterministic regressors,
1.: curs when the trend is linear. such as an intercept term or a deterministic time crend. As such, there is a so-
phisticated set of procedures that can aid in the identification process. These
The benefit of the Hodrick-Prescott decomposition is that it can extract the same
¡¡:: trend from a set of variables. For example, many real business cycle models indi- procedures can be used if it is not known what deterministic elements are part of
l
cate that ali variables will have the same stochastic trend. A Beveridge and Nelson the true data-generating process. It is important to be wary of the results from
1
l such tests since (1) they ali ha ve low power to discriminate between a unit roo1
¡ decomposition separately applied to each variable will not yield the same trend for
each.
and near unit root process and (2) they may have used an inappropriate set of de-
terministic regressors.
:.·:' .
 UNIT ROOT PROCESSES Thus, if a 1 = l, the variance becomes infinitely large as t increases. Under the
null hypothesis, it is inappropriate to use classical statistical methods to estimate
As shown in the last chapter, there are important differences between stationary and and perfonn significance tests on the coefficient a 1 • If the {y,} sequence is gener-
nonstationary time series. Shocks to a stationary time series are necessarily tempo- ated as in (4.2), it is simple to show that the OLS estimate of (4.1) will yield a bi-
rary; over time, the effects of the shocks will dissipate and the series will revert to ased estimate of a 1 • In Section 8 of the previous chapter, it was shown that the first-
its long-run mean leve!. As such, long-terrn forecasts of a stationary series will con- order autocorrelation coefficient in a random walk model is
verge to the unconditional mean of the series. To aid in identification, we know that
p 1 = [(t - l)lt] 05 < 1

l
a covariance stationary series:
l. Exhibits mean reversion in that it fluctuates around a constant long-run mean. Since the estimate of a 1 is directly related to the value of p 1 , the estimated value
2. Has a finite variance that is time-invariant. of a 1 is biased to be below its true value of unity. The estimated model will mimic
that of a stationary AR( 1) process with a near unit root. Hence, the usual t-test can- "
3. Has a theoretical correlogram that diminishes as lag length increases.
not be used to test the hypothesis a 1 = 1.
On the other hand, a nonstationary series necessarily has pennanent components. Figure 4.1 shows the sarnple correlograrn for a simulated random walk process.
The mean andJor variance of a nonstationary series are time-dependent. To aid in One hundred norrnally distributed random deviates were obtained so as to mimic
the identification of a nonstationary series, we know that: the {E,) sequence. Assuming y 0 =O, we can calculate the next 100 values in the (y,}
1. There is no long-run mean to which the series retums. sequence as y,= y,_, + E1• This particular correlogram is characteristic of most sam-
2. The variance is time-dependen! and goes to infinity as time approaches infinity. ple correlograms constructed from nonstationary data. The estimated value of p 1 is
close to unity and the sample autocorrelations die out slowly. If we did not know
3. Theoretical autocorrelations do not decay but, in finite samples, the sample cor-
the way in which the data were generated, inspection of Figure 4.1 might lead us to
relogram dies out slowly.
falsely conclude that the data were generated from a stationary process. With this
. · Although the properties of a sample correlogram are useful tools for detecting particular data, esti mates of an AR( 1) model with and without an intercept yield
the possible presence of unit roots, the method is necessarily irnprecise. What may (standard errors are in parentheses):
appear as a unit root to one observer may appear as a stationary process to another.
The problem is difficult because a near unit root process will have the same shaped y,::: 0.9546yt-1 +E,, R 2 = 0.86 (4.3)
ACF as a unit root process. For example, the correlogram of a stationary AR(l) (0.030)
process such that p(l) = 0.99 will exhibit the type of gradual decay indicative of a y,::: 0.164 + 0.9247y,_I +E,, R 2 = 0.864 (4.4)
nonstationary process. To illustrate sorne of the issues involved, suppose that we (0.037)
know a series is generated from the following first-order process: 1
Examining (4.3), a careful researcher would not be willing to dismiss the possi-
y,::: ª1Y1-1 +E, (4.1) bility of a unit root since the estimated value of a 1 is only 1.5133 standard devia-
tions from unity. We might correctly recognize that under the null hypothesis of a
where (E,} is generated frorn a white-noise process. unit root, the estimate of a 1 will be biased below unity. If we knew the true distrib-
First, suppose that we wish to test the null hypothesis that a 1 = O. Under the ution of a 1 under the null of a unit root, we could perfonn such a significance test.
maintained null hypothesis of a 1 =O, we can estimate (4.1) using OLS. The fact Of course, if we did not know the true data-generating process, we might estimate
that E, is a white-noise process and \a 1 1 < 1 guarantees that the [y,} sequence is the model with an intercept. In (4.4), the estímate of a, is more than two standard
stationary and the estimate of a 1 is efficient. Calculating the standard error of the deviations from unity: (1 - 0.9247)/0.037 = 2.035. However, it would be wrong to
estimate of a 1 , the researcher can use a r-test to detennine whether a 1 is signifi- use this infonnation to reject the null of a unit root. After ali, the point of this sec-
cantly different from zero. tion has been to indicate that such t-tests are inappropriate under the null of a unit ·
The situation is quite different if we want to test the hypothesis a 1 = l. Now, un- roo t.
der the null hypothesis, the {y,} sequence is generated by the nonstationary process: Fortunately, Dickey and Fuller (1979, 1981) devised a procedure to fonnally test
, for the presence of a unit root. Their methodology is similar to that used in con-
structing the data reported in Figure 4.1. Suppose that we generated thousands of
y,::: LE¡ (4.2) such random walk sequences and for each we calculated the estimated value of a 1 •
i=l Although most ali of the estimates would be close to unity, sorne would be further
 Figure 4.1 The application of these Dickey-Fuller critica/ values to tests for unit roots is
/
/-
A simulated random walk process. straightforward. Suppose we did not know the true-data generating process and
/
10
were trying to ascertain whether the data used in Figure 4.1 contained a unit root.
Using these Dickey-Fuller statistics, we would not reject the null of a unit root in
8
(4.4). The estimated value of a 1 is only 2.035 standard deviations from unity. In
6 fact, if the true value of a 1 does equal unity, we should find the estimated value to
be within 2.58 standard deviations from unity 90% of the time.
4 Be aware that stationarity necessitates -1 < a 1 < l. Thus, if the estimated value
of a 1 is close to -1, you should also be concerned about nonstationarity. If we de-
2 fine y :::: a 1 - 1, the equivalent restriction is -2 < y < O. In conducting a Dickey-
3
Fuller test, it is possible to check that the estimate1 value of y is greater than -2.
o
Monte Cario Simulation
-2
The procedure Dickey and Fuller (1979, 1981) used to obtain their critica! values is
-4 typical of that found in the modern time series literature. Hypothesis tests concem-
o
ing the coefficients of non-stationary variables cannot be conducted using tradi-
tional t-tests or F-tests. The distributions of the appropriate test statistics are non-
standard and cannot be analytically evaluated. However, given the trivial cost of
Correlogram of the process. computer time, the non-standard distributions can easily be derived using a Monte
.. Cario simulation .
The first step in the procedure is to computer generate a set of random numbers
0.8 (sometimes called pseudo-random numbers) from a given distribution. Of course,
the numbers cannot be entirely random since ali computer algorithms rely on a de-
terministic number generating mechanism. However, the numbers are drawn so as
0.6
to mimic a random process having sorne specified distribution. Usually, the num-
bers are designed to be nonnally distributed and serially uncorrelated. The idea is to
0.4 use these numbers to represent one replication of the entire {e,} sequence.
Ali major statistical packages have a built-in random number generator. An inter-
0.2
esting experiment is to use your software package to draw a set of 100 random
numbers and check for serial correlation. In almost ali circumstances, they will be
highly correlated. In your own work, if you need to use serially uncorrelated num-
o 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 bers, you can model the computer generated numbers using the Box Jenkins
methodology. The residuals should approximate white noise.
(b)
The second step is to specify the parameters and initial conditions of the {y,} se-
quence. Using these parameters, initial conditions, and random numbers, the {y,}
from unity than others. In performing this experiment, Dickey and Fuller found that can be constructed. Note that the simulated ARCH processes in Figure 3.9 and ran-
in the presence of an intercept: dom-walk process in Figure 4.1 were constructed in precisely this fashion.
Similarly, Dickey and Fuller (1979, 1981) obtained 100 values for {E,}. set a 1 == 1,
Ninety percent of the estimated val ues of a 1 are lcss than 2.5 8 standard errors from
y 0 ==O, and calculated l 00 values for {y,} accordíng to (4.1 ). At thís point, the para-
unity.
meters of interest (such as the estímate of a 1 or the in-sample variance of y,) can be
Ninety-five percent of the estimated values of a 1 are less than 2.89 standard errors obtained.
from unity. The beauty of the method is that all important attributes of the constructed {y,}
Ninety-nine percent of the estimated values of a 1 are less than 3.51 standard errors sequence are known to the researcher. For this reason, a Monte Cario simulation is
from unity. 2 often referred to as an "experiment." The only problem is that the set of random
 • .. _..,,.Ó )'-" • ' - ,...,._, ...,,.... '- "•• • •v ·

In their Monte Cario analysis, Granger and Newbold generated many such sam-
numbers drawn is just one possible outcome. Obviously, the estimates in (4.3) and
(4.4) are dependent on the values of the simulated { E1 } sequence. Different out- ples and for each sample. estimated a regression in the form of (4.5). Since the (y 1}
comes for { E1 } will yield different values of the simulated (y 1 } sequence. and {.zrl sequences ar md pendent of each other, Equation (4.5) is necessarily
This is why the Monte Cario studies perform many replications of the process meamngles ; a y relat10nsh1p between the two variables is spurious. Surprisingly,
outlined above. The third step is to replicate steps 1 and 2 thousands of times. The at the 5% s1g01ficance leve.!, they were able to reject the null hypothesis a 1 = O in
goal is to ensure that the statistical properties of the constructed (y, l sequence are approx1mately 750:º of the t1 e. Moreover, the regressions usually had very high R 2
in accord with the true distribution. Thus, for each replication, the parameters of in- va!ues and the est1mated res1duals exhibited a high degree of autocorrelation.
terest are tabulated and critica! values (or confidence intervals) obtained. As such, To xplain the ?ranger a d Newbold findings, note that the regression equation
the properties of your data set can be compared to the properties of the simulated (4.5! is ne essanly meanmgless if the residual series {e,} is nonstationary.
data so that hypothesis tests can be performed. This is the justification for using the Obv1ously, 1f the {e uence has a stochastic trend, any error in period t never
Dickey-Fuller critica! values to test the hypothesis a 1 = 1. dec ys, so th.at tne deviation from the mOcieCíSpei:1nanerit. It is hard to imagine at-
tachmg any 1mportance to an economic model having perrnanent errors. The sim-
One limitation of a Monte Cario experiment is that it is specific to the assump-
plest way to examine the properties of the {e,) sequence is to abstract from the in-
tions used to generate the simulated data. If you change the sample size, include (or
tercept term a 0 and rewrite (4.5) as
delete) an additional parameter in the data generating process, or use altemative ini-
tial conditions an entirely new simulation needs to be performed. Nevertheless, you
shou!d be able to envision many applications of Monte Cario simulations. As dis-
cussed in Hendry, Neale, and Ericsson (1990), they are particularly use ful for
studying the small sample properties of time-series data. As you will see shortly, lf z, and Y1 are id by (4.6) and (4.7), we can impose the initial eortdi:tions
Monte Cario simulations are the workhorse of unit root tests. Yo = X 0 =O, so that:

Unit Roots in a Regression Model

The unit root issue arises quite naturally in the context of the standard regression
4
(4.8)
model. Consider the regression equation:
Clearly, the variance of the error becomes infinitely Iarge as t increases. More-
Y1 = ªº + ª1Z1 + e1 (45)
over, the error has a permanent component in that E,e 1+ 1 = e1 for ali ¡ O. Hence, the
assumption; embedded in th usual ypothesis tests are violated, so that any t-test,
The assumptions of the classical regression model necessitate that both the {y,}
F-test, or R values are unreliable. lt 1s easy to see why the estimated residuals from
and ( Z1} sequences be stationary and the errors have a zero mean and finite vari-
- pu ous regression will exhibi _i high_9 gree Lau_toc]5-fyelition. Updating ( 4.8),
ance. In the presence of nonstationary variables, there might be what Granger and
you sfiould be able to demonstrate that the theoretical value of the correlation coef-
Newbold (1974) call a spurious regression. A spurious regression has a high R 2 , ficient between e1 and e,+ 1 goes to unity as t increases.
t-statistics that appear to be significant, but the results are without any economic
meaning. The regression output "looks good" because the least-squares estimates . The essence of the problem is that if a 1 =O, the data generating process in (4.5)
is Y1 = a0 + Er Given that {y1 ) is integrated of order one [i.e., 1(1)], it follows that
are not consistent and the customary tests of statistical inference do not hold.
( e 1 ) is I( 1) under the null hypothesis. However, the assumption that the error term
Granger and Newbold (1974) pro vide a detailed examination of the consequences
is a unit root pr cess is incon.sistent with the distributional theory underlying the
of violating the stationarity assumption by generating two sequences, (y1} and (z1},
use of OLS. Th1s problem w1!1 not disappear in large samples. In fact, Phillips
as independent random walks using the formulas:
( 1986) proves that the larger the sample, the more likely you are to falsely conclude
that a 1 t:- O.
Y1 = Y1-I + Ey 1
(4.6)
Worksheet 4.1 illustrates the problem of spurious regressions. The top two
graphs show 100 realizations of the {y,} and ( z,} sequences generated according to
and
(4.6) and (4.7). Although {Ey1 ) and {Ezrl are drawn from white-noise distributions
Z1 = Z1-I +Ea (4 7) the realizations of the two sequences are such that y 100 is positive and z 100 negative'.
You can see that the regression of y, on z, captures the within-sample tendency of
the sequences to move in opposite directions. The straight line shown in the scatter
where Eyr and et,= whhe-noise processes independent of each other
 plot is the OLS regression line y, = --0.31 - 0.46z,. The correlation coefficient be- the errors (i.e., u,) will pull the relationship further and further from -2.0. The
tween (y,} and ( z,} is --0.372. The residuals from this regression have a unit root; scatter plot of the two sequences suggests that the R 2 statistic will be close to unity;
/ as such, the coefficients --0.31 and --0.46 are spurious. Worksheet 4.2 illustrates the in fact, R 2 is almost 0.97. However, as you can see in the last graph of Worksheet
same problem using two simulated random walk plus drift sequences: y,= 0.2 + y,_ 1 4.2, the residuals from the regression equation are nonstationary. Ali departures
+ Ey, and z, = --0.1 + z,_ 1 + Ezr. The drift terms domínate, so that for small values of t, from this relationship are necessarily perrnanent.
it appears that y, = -2z,. As sample size increases, however, the cumulated sum of The point is that the econometrician has to be very careful in working with non-
stationary variables. In terrns of (4.5), there are four cases to consider:

WORKSHEET 4.1 Spurious Regressions: Example 1 ' CASE 1

Consider the two random walk processes:

Both (y,) and (z,) are stationary. When both variables are stationary, the classical
z,=z,_,+E,, regression model is appropriate.

CASE2
o o
The {y,} and {z,} sequences are integrated of different orders. Regression equations
y, z, using such variables are meanini?iless. For exampl , replace (4.7) by the stationary
-2 -2 process z, = pz,_ 1 + Ezr• where 1 p \ < 1. Now (4.8) IS replaced by e,= LEy; - L:p'Ezr_,.
Although the expression Lp;E,,_; is convergent, the ( Ey,} sequence still contains a
trend component. 5
-4 -4
o 50 100 o 50 100
CASE 3

Since the ( Eyr} and { Ezr} sequences are independent, the regression of y, on z, is spu- The nonstationary {y,} and {z,} sequences are integrated of the same order and the
rious. Given the realizations of the random disturbances, it appears as if the two se- residual sequence contains a stochastic trend. This is the case in which the regres-
quences are related. In the scatter plot of y, against z,, you can see that y, tends to sion is spurious. The results from such spurious regressio!ls Wt:. meaningless in that
rise as z, decreases. The regression equation of y, on z, will capture this tendency. errors_a.:r!!J> !!ll -In thls case, Ttls often-rec;i'luñendedthat-th·é-reg;:; sio
The correlation coefficient between y, and z, is --0.372 and a linear regression yields equation be estimated in first differences. Consider the first difference of (4.5):
y, = --0.46z, - 0.31. However, the residuals from the regression equation are nonsta-
tionary.

Scatter plot of y, and z, Regression residuals Since y,, z,, and e, each contain unit roots, the first difference of each is stationary.
2 2 Hence, the usual asymptotic results apply. Of course, if one of the trends is deter-
++ + +
+++ ++/ ministic and the other is stochastic, first-differencing each is not appropriate.
o + +
+ i: .t + + o CASE4
.t
...
+ +
+.t+ :e
-2 1.+ ... -2 The nonstationary {y,} and { z,} sequences are integrated of the same order and the
+ residual sequence i station ry. In this circumstance, {y,} and (z,} are cointe-
grated. A trivial example of a coiritegrated system occurs if Ezr and Ey, are perfectly
-4 -4
-3 -2 -1 o 2 o 50 100 correlated. If Ezr = Ey,, then (4.8) can be set equal to zero (which is stationary) by
+
Scatter plot
setting a, = l. To considera more interesting example, suppose that both z, and y,
-Regression line are the random walk plus noise processes:
 y,=µ,+ Eyt Cases 2 or 3 apply. If the variables are cointegrated, the results of Chapter 6 apply.
Z, = µ, + E, 1 The rcmainder of this chapter considers the formal test procedures for the presence
of unit roots and/or deterministic time trends.
where Ey, and E" are white-noise processes and µ, is the random walk process µ, =
µ,_ 1 + Er Note that both {z,) and {y,) are unit root processes, but y, - z, = Ey, - E,, is
stationary.
Ali of Chapter 6 is devoted to the issue of cointegrated variables. For now, it is 2. DICKEY-FULLER TESTS
sufficient to note that pretesting the variables in a regression for nonstationarity is
extremely importan t. Estimating a regression in the form of (4.5) is meaningless if The last section outlíned a simple procedure to detennine whether a 1 = l in the
model y,= a 1 y,_ 1 +E,. Begin by subtracting y,_ 1 from each side of the equation in
order to write the equivalent form: !::.y,= yy,_ 1 +E" where y= a 1 - l. Of course, test-
WORKSHEET 4.2 Spurious Regressions: Example 2 ing the hypothesis a 1 = 1 is equivalent to testing the hypothesis y= O. Dickey and
Fuller ( 1979) actually consider three different regression equations that can be u sed
Consider the two random walk plus drift processes: to test for the presence of a unit root:

y,= 0.2 + y,_ 1 + Ey 1 z,= -0.1 + z,_, +E" !::.yr == YY1-1 +e., (4.9)
30 o t.yr ==ªo+ 'YYr-1 + e., (4.10)

20
t.yr == ªº + YY1-1 + a2t +e., (4.l l)

-5 --
The difference between the three regressions concems the presence of the deter-
10 z, ministic elements a 0 and a 2 t. The first is a pure random walk model, the second
-10 adds an intercept or drift term, and the third includes both a drift and linear time
o tren d.
The parameter of interest in ali the regression equations is y, if y= O, the {y1 } se-
-10 -150 quence contains a unit root. The test involves estimating one (or more) of the equa-
o 50 100 50 100
tions above using OLS in order to obtain the estimated value of y and associated
standard error. Comparing the resulting t-statistic with the appropriate value re-
ported in the Dickey-Fuller tables allows the researcher to determine whether to ac-
Again, the {e.Y,) and {E."} sequences are independent, so that the regression of y, on
cept or reject the null hypothesis y= O.
z, is spurious. The scatter plot of y, against z, strongly suggests that the two series
are related. It is the deterministic time trend that causes the sustained increase in y, Recall that in (4.3), the estímate of y,= a 1 y,_ 1 +e., was such that a 1 = 0.9546 with
and sustained decrease in z,. The residuals from the regression equation y, = -2z, + a standard error of 0.030. Clearly, the OLS regression in the form !::.y,= yy,_ 1 + e.,
will yield an estímate of y equal to --0.0454 with the same standard error of 0.030.
e, are nonstationary.
Hence, the associated r-statistic for the hypothesis y= O is -1.5133 (-0.0454/0.03 =
-1.5133).
Scatter plot of y, and z, Regression residuals
4 The methodology is precisely the same, regardless of which of the three forms of
the equations is estimated. However, be aware that the critica! values of the r-statis-
tics do depend on whether an intercept and/or time trend is included in the regres-
20 2
sion equation. In their Monte Cario study, Dickey and Fuller (1979) found that the
y, 10 y,+ 2·z, o critica! values for y= O depend on the form of the regression and sample size. The
statistics labeled -r, -rµ, and -r, are the appropriate statistics to use for Equations
(4.9), (4.10), and (4.11), respectively.
o -2
Now, look at Table A at the end of this book. With 100 observations, there are
-10L---'................. -'- J three different critica! values for the t-statistic y= O. For a regression without the
-15 -10 -5 o 50 100 intercept and trend terms (a 0 = a2 = 0), use the section labeled -r. With 100 observa-
z, tions, the critica! values for the r-statistic are -1.61, -1.95 and -2.60 at the 10, 5,
 /
,,and 1 % significance levels, respectively. Thus, in the hypothetical example with y= r = number of restrictions
-0.0454 and a standard error of 0.03 (so that t = -1.5133), it is not possible to reject T = number of usable observations
the null of a unit root at conventional significance levels. Note that the appropriate k = number of parameters estimated in the unrestricted model
critica] values depend on sample size. As in most hypothesis tests, for any given
leve] of significance, the critica! values of the t-statistic decrease as sample size in- Hence, T - k = degrees of freedom in the unrestricted model.
creases. Comparing the calculated value of <ji¡ to the appropriate value reported in Dickey
Including an intercept term but nota trend term (only a 2 =O) necessitates the use and Fuller ( 1981) allows you to determine the significance leve! at which the re-
of the critica! values in the section labeled 'tµ- Estimating (4.4) in the form ti.y, = striction is binding. The null hypothesis is that the data are generated by the re-
ªº + YY1-1 +E, necessarily yields a value of y equal to (0.9247 - 1) = -0.0753 with a stricted model and the altemative hypothesis is that the data are generated by the
standard error of 0.037. The appropriate calculation for the 'tµ statistic yields unrestricted model. If the restriction is not binding, RSS(restricted) should be close
-0.0753/0.037 = -2.035. If we read from the appropriate row of Table A, with the to RSS(unrestricted) and <ji¡ should be small; hence, large values of <jl¡ suggest a
same 100 observations, the critica] values are -2.58, -2.89, and -3.51 at the 1O, 5, binding restriction and rejection of the null hypothesis. Thus, if the calculated value
and 1 % significance levels, respectively. Again, the null of a unit root cannot be re- of <ji, is smaller than that reported by Dickey and Fuller, you can accept the re-
jected at conventional significance levels. Finally, with both intercept and trend, stricted model (i.e., you do not reject the null hypothesis that the restriction is not
use the critica] values in the section labeled 't,; now the critica! values are -3.45 binding). If the calculated value of <)l; is larger than reported by Dickey and Fuller,
and -4.04 at the 5 and 1 % significance levels, respectively. The equation was not you can reject the null hypothesis and conclude that the restriction is binding. The
estimated using a time trend; inspection of Figure 4.1 indicates there is little reason critica] values of the three <)l; statistics are reported in Table C at the end of this text.
to include a deterministic trend in the estimating equation. Finally, it is possible to test hypotheses conceming the significance of the drift
As discussed in the next section, these critica! values are unchanged if (4.9), term a0 and time trend a 2 . Under the null hypothesis y= O, the test for the presence
(4.10), and (4.11) are replaced by the autoregressive processes: 6 of the time trend in (4.14) is gíven by the 't¡h statistic. Thus, this statistic is the test
a 2 =O given that y= O. To test the hypothesis a 0 =O, use the 'ª' statistic if you esti-
p

ti.y, = 'YYr-1 + L ¡li.Yr-i+I +E, (4 12)

mate (4.14) and the 'taµ statistic if you estimare (4.13). The complete set of test sta-
tistics and their critica! values for a sample síze of 100 are sumrnarized in Table
i=2
p
4.1.
t..yr =ªo +yy,_¡ + L ;li.Yr-i+I +Er (4.13)
i=2
p
' Table 4.1 Summary of the Dickey-Fuller Tests
t:i.yr =ªo+ 'YYr-1 + a2t + L ;ÓYr-i+I +E, (4.14)
i=2 Critica! values for
The same t, 'tµ, and t, statistics are ali used to test the hypotheses y= O. Dickey 9?% and 99%
Model Hypothesis Test Statistic Confidence Intervals
and Fuller (1981) provide three additional F-statistics (called <jl 1 , <j> 2 and <jl 3) to test
joint hypotheses on the coefficients. With (4.10) or (4.13), the null hypothesis y= y=O t, -3.45 and - 4.04
ªº::::O is tested using the <jl 1 statistic. Including a time trend in the regressíon-so a 0 =O given y= O 3.11 and 3.78
'ta<
that (4.11) or (4.14) is estimated-the joint hypothesis a 0 = y= a 2 = O is tested us-
ing the <!> 2 statistic and the joint hypothesis y= a 2 = ú is tested using the <j> 3 statistic. a 2 =O given y= O i: , 2.79 and 3.53
The <!>1, <Pi. and <jl 3 statistics are constructed in exactly the same way as ordinary y= ª2 =o <jl3 6.49 and 8.73
F-tests are: ªº ="(= ª2 =o <1>2 4.88 and 6.50
\'
D.y, =ªo+ 'YY1-1 +E, y=O 'tµ -2.89 and -3.51
<P = [RSS(restricted)- RSS(unrestricted)]/r
a 0 =0giveny=O 'taµ 2.54 and 3.22
' RSS(unrestricted)/(T- k)
ao=r=O <P 1 4.71 and 6.70
ó.y, = "fY1-I + E, y=O 't -1.95 and -2.60
where RSS(restricted) and RSS(unrestricted) = the sums of the squared residuals
from the restricted and unrestricted models Nores: Critica! values are for a sample size of 1OO.
An Example 3. EXTENSIONS OF THE DICKEY-FULLER TEST
To illustrate the use of the various test statistics, Dickey and Fuller ( 1981) use quar-
terly values of the Jogarithm of the Federal Reserve Board's Production Index over Not ali time-series processes can be well represented by the first-order autoregres-
sive process tiy, = a 0 + yy,_ 1 + a 2 t +E,. It is possible to use the Dickey-Fuller tests
the 1950:1to1977:IV period to estimate the following three equations:
in higher-order equations such as (4.12), (4.13), and (4.14). Consider the pth-order
auroregressive process:
tiy, = 0.52 + O.CXl 120! - O. l l 9y,_ I + 0.498tiy,_ I + E" RSS = 0.056448
(0.15) (0.00034) (0.033) (0.081) (4.15)
lly1 = 0.0054 + 0.447f>y,_I +E,, RSS = 0.063211
(0.0025) (0.083) (4.16)
tiy, = 0.51 ltiyt-l +E,, RSS = 0.065966 To best understand the methodology of the augmented Dickey-Fuller test, add
(0.079) (4.17)
and subtract apYr-p+I to obtain:

where RSS =residual sum of squares, and standard errors are in parentheses.
To test the null hypothesis that the data are generated by (4.17) against the alter-
native that (4.15) is the "true" model, use the qi 2 statistíc. Dickey and Fuller test the Next, add and subtract (ap-- 1 + aP)y,_r+ 2 to obtain
null hypothesis a0 = a 2 =y= O as follows. Note that the residual sums of squares of
the restricted and unrestricted models are 0.065966 and 0.056448 and the null hy-
pothesis entails three restrictions. With l JO usable observations and four estimated
parameters, the unrestricted model contains 106 degrees of freedom. Sin ce Continuing in this fashion, we get
0.056448/106 = 0.000533, the (j> 2 statistic is given by
p

qi2 = (0.065966 - 0.056448)/ 3(0.000533) = 5.95 6.yr =ªo+ 'Y Yr-1 + L ¡LlYr-i+I + E1
i=2

With l JO observations, the critica! value of (j> 2 calculated by Dickey and Fuller is
5.59 at the 2.5% significance leve!. Hence, it is possible to reject the null hypothe-
where 'Y=-(!- f., a;)
r=I
sis of a random walk against the altemative that the data contain an intercept and/or p
a unit root and/or a deterministic time trend (i.e., rejecting a0 = a 2 = y= O means
that one or more of these parameters does not equal zero).
¡= Lª1
j=i
(4.19)

Dickey and Fuller also test the null hypothesis a2 =y= O given the altemative of
(4.15). Now ifwe view (4.16) as the restricted model and (4.15) as the unrestricted In (4.19), che coefficient of interest is y; if y= O, the equation is entirely in first
model, the <j> 3 statistic is calculated as differences and so has a unit root. We can test for the p esence of a unit root using
the same Dickey-Fuller statistics discussed above. Again, the appropriate statistic
<j>3 = (0.063211 - 0.056448)/ 2(0.000533) = 6.34 to use depends on the deterministic components included in the regression equa-
tion. Without an intercept or trend, use the 1statistic; with only the intercept, use
With 11O observations, the critica! value of <j> 3 is 6.49 at the 5% significance the 1:µ statistic; and with both an intercept and trend, use the e, statistic. It is worth-
leve! and 5.47 at the 10% significance level. 7 At the 10% leve!, they reject the null while pointing out that the results here are perfectly consistent with our study of
hypothesis. However, at the 5% leve!, the calculated value of <\> 3 is smaller than the difference equations in Chapter 1. If the coefficients of a difference equation sum to
critica! value; they do not reject the null hypothesis that the data contain a unit root 1, ar least one characteristic root is unity. Here, if Úl; = 1, y= O and the system has
and/or deterministic time trend. a unit root.
To compare with the 1, test (i.e., the hypothesis that only y= 0) note that Note that the Dickey-Fuller tests assume that the errors are independent and
ha ve a constant variance. This raises four important problems related to the fact that
1, = --0.119/0.033 =-3.61. we do not know the true data-generating process. First, the true data-generating
process may contain both autoregressive and moving average components. We
which rejects the null of a unit root at the 5% leve!. need to know how to conduct the test if the order of the moving average terms (if
 unknown. Second, we cannot properly estímate y and its standard error un- increased number of lags necessitates the estimation of additional parameters and a
/ less ali the autoregressive terms are included in the estimating equation. Clearly, loss of degrees of freedom. The degrees of freedom decrease since the number of
the simple regression 6y, = a 0 + yy,_ 1 + E, is inadequate to this task if (4.18) is the parameters estimated has increased and because the number of usable observations
'
true data-generating process. However, the true order of the autoregressive process has decreased. (We lose one observation for each additional lag included in the au- j
is usually unknown to the researcher, so that the problem is to select the appropriate toregression.) On the other hand, too few lags will not appropriately capture the ac-

1
lag length. The third problem stems from the fact that the Dickey-Fuller test con- tual error process, so that y and its standard error will not be well estimated.
siders only a single unit root. However, a pth-order autoregression has p character- How <loes the researcher select the appropriate lag length in such circumstances'l
istic roots; if there are m :S: p unit roots, the series needs to be differenced m times to One approach is to start with a relatively long lag length and pare down the model
achieve stationarity. The fourth problem is that it may not be known whether an in- by the usual t-test and/or F-tests. For example, one could estímate Equation (4.20)
tercept and/or time trend belongs in (4.18). We consider the first three problems be- using a lag length of n*. If the t-statistic on lag n* is insignificant at sorne specified
low. Section 7 is con cerned with the issue of the appropriate deterministic regres- critica! value, reestimate the regression using a lag length of n* - 1. Repeat the
sors. process until the lag is significantly different from zero. In the pure autoregressive
Since an invertible MA model can be transformed into an autoregressive model, case, such a procedure will yield the true lag length with an asymptotic probability
the procedure can be generalized to allow for moving average components. Let the of unity, provided that the initial choice of lag length includes the true length. With
{y,} sequence be generated from the mixed autoregressive/moving average process: seasonal data, the process is a bit different. For example, using quarterly data, one
could start with 3 years of lags (n = 12). If the t-statistic on lag 12 is insignificant at
A(L)y, = C(L)E, sorne specified critica! value and an F-test indicates that lags 9 to 12 are also in-
significant, move to lags 1 to 8. Repeat the process for lag 8 and lags S to 8 until a
where A(L) and C(L) = polynomials of orders p and q, respectively reasonable lag length has been determined.
Once a tentative lag length has been determined, diagnostic checking should be
If the roots of C(L) are outside the unit circle, we can write the {y,} sequence as
conducted. As always, plotting the residuals is a most important diagnostic too!.
the autoregressive process: There should not appear to be any strong evidence of structural change or serial
correlation. Moreover, the correlogram of the residuals should appear to be white
A(L)y, /C(L) = E, noise. The Ljung-Box Q-statistic should not revea! any significant autocorrelations
among the residuals. It is inadvisable to use the altemative procedure of beginning
or, defining D(l) = A(l)IC(L), we get with the most parsimonious model and keep adding lags until a significant lag is
found. In Monte Cario studies, this procedure is biased toward selecting a value of
D(L)y, =E,
n that is less than the true value.
Even though D(l) will generally be an infinite-order polynomial, in principie we
Multiple Roots
can use the same technique as used to obtain (4.19) to form the infinite-order au-
toregressi ve model: Dickey and Pantula (1987) suggest a simple extension of the basic procedure if
more than one unit root is suspected. In essence, the methodology entails nothing
more than performing Dickey-Fuller tests on successive differences of (y,}. When
t,.y, = 'YY1-l + IJl;6Y,-i+l +E, (4.20) exactly one root is suspected, the Dickey-Fuller procedure is to estímate an equa-
i=2 tion such as 6y, = a0 + yy,_ 1 + E,. Instead, if two roots are suspected, estímate the
equation:
As it stands, (4.20) is an infinite-order autoregression that cannot be estimated Now, the second problem conceming the appropriate lag length arises. Including
using a finite data set. Fortunately, Said and Dickey (1984) have shown that an un- too many lags reduces the power of the test to reject the null of a unitroot since the
known ARIMA(p, 1, q) process can be well approximated by an ARIMA(n, 1, O)
autoregression of order no more than T 113 • Thus, we can solve the first problem by
using a finite-order autoregression to approximate (4.20). The test for y= O can be
conducted using the aforementioned Dickey-Fuller 1, "tµ, or 1, test statistics.
 (4.21)

Use the appropriate statistic (i.e., 1, 'µ• or 1:, depending on the deterministic
ele- ments actually included in the regression) to determine whether 1 is
significantly different from zero. If you cannot reject the null hypothesis that
1 =O, conclude
that the {y,} sequence is /(2). If 1 does differ from zero, go on to determine
whether there is a single unit root by estimating
 --------------------------------------------·- •., .1 ............. ''-"'-.!
- - -- -
...................

(4 22) Tourist visits to Spain have a unit root and seasonal unit root. Since seasonality
is a key feature of many economic series, a sizable literature has developed to test
Since there are not two unit roots, you should find that P1 and/or P2 differ from for seasonal unit roots. Befare proceeding, note that the first difference of a sea-
zero. Under the null hypothesis of a single unit root, P1 <O and P2 =O; un er the al- sonal unit root process will not be stationary. To keep matters simple, suppose that
temative hypothesis, (y }1 is stationary, so that P 1 and p2 are both negat1ve. Thus, the quarterly observations of (y1 ) are generated by
estimate (4.22) and use the Dickey-Fuller critica! values to test the null hypothes1s
p 1 =O. If you reject this null hypothesis, conclude that {y 1).is stationary. Yr = Yr-4 + Er
As a rule of thumb, economic series do not need to be d1fferenced more than two
times. However, in the odd case in which at most r unit roots are suspected, the Here, the fourth difference of {y,} is stationary; using the notation of Chapter 2,
procedure is to first estímate we can write 64,}', = E1. Given the initial condition y 0 = y_ 1 = ... = O, the solution for
Ar (:t Ar-]
y, is:
Ll Yr =ªo+ 1-'1Ll Yr-1 +E,

If !J..ry, is stationary, you should find that -2 < 1 <O. If the Dickey:-Fuller critica] y,= E1 + E 1-4 + E 1_ 8 + ...
values for p 1 are such that it is not possible to reject the nul of a ntt root, you ac-
cept the hypothesis that (y,} contains r unit roots. If you reJect th1s null of exactly so that
r unit roots, the next step is to test for r - 1 roots by estimating
1/4 1/4

Y1 - Y1-1 = L
i=O
E4; - L
i=O
E4i-l

If both p 1 and p2 differ from zero, reject the null hypothesis of r - 1 u.nit root .
You can use the Dickey-Fuller statistics to test the null of exactly r - 1 unlt roots if Hence, Ay, equals the difference between two stochastic trends. Since the vari-
the t statistics for p 1 and p2 are both statistically different from zero. If you can re- ance of !J..y1 increases without Iimit as t increases, the (Liy1} sequence is not station-
ject this null, the next step is to form ary. However, the seasonal difference of a unit root process may be stationary. For
¡.,,ry, = ao +
(:t Ar-!
+ (:t F-2
+ (:t Ar-3 +E example, if {y1} is generated by y 1 = y 1_ 1 + E 1, the fourth difference (i.e., Ll4,}' 1 = E 1 +
Yi-1 Yr-1 Yr-1 r E,_ 1 + E,_ 2 + E,_ ) is stationary. However, the variance of the fourth difference is
J-l¡Ll l-'2Ll l-'3Ll 3
larger than the variance of the first difference. The point is that the Dickey-Fuller
As long as it is possible to reject the null hypothesis that the variou& values of the procedure must be modified in order to test for seasonal unit roots and distinguish
p; are nonzero, continue toward the equation: between seasonal versus nonseasonal roots.
There are severa! alternative ways to treat seasonality in a nonstationary se-
quence. The most direct method occurs when the seasonal pattern is purely deter-
ministic. For example, Jet D 1 , D2 , and D 3 represent quarterly seasonal dummy vari-
Continue in this fashion until it is not possible to reject the null of a unit root or ables such that the value of D; is unity in season iand zero otherwise. Estímate the
the y 1 series is shown to be stationary. Notice that this procedure i.s quite diffe:ent regression equation:
from the sequential testing for successively greater numbers of untt ro.ots. It m1ght
seem tempting to test for a single unit root and, if the null cannot be. reJected, go on
(4.23)
to test for the presence of a second unit root. In repeated samples, th1s method tends
to select too few roots.
where y, is the regression residual, so that y 1 can be viewed as the deseasonalized
Seasonal Unit Roots value of y 1•
Next, use these regression residuals to estímate the regression:
You will recall that the best-fitting model for the monthly Spanish tourism data
used in Chapter 2 had the form: p

12 LÍY1 ='Y Y1-I + Ll3¡LlY1-i+I +El

(1 - L 12 )(1 - L)yl = (1 + P1L)(l + P12L )E1 i=2
he null hypothesis of a unit root (i.e., y = 0) can be tested using the Dickey- CASE 3
Fuller 'tµ statistic. Rejecting the null hypothesis is equivalent to accepting the alter-
native that the {y,} sequence is stationary. The test is possible as Dickey, Bell, and
lf ei her a 3 or a4 is equal to unity, the {y,} sequence has an annual cycle. For exam-
Miller ( 1986) show that the limiting distribution for y is not affected by the removal
of the deterministic seasonal components. If you want to include a time trend in ple, f. 3_=_ 1, a ho o eneous solut n to (4.25) is y,= iy,_ 1 • Thus, if y,= 1, Yr+i = i,
(4.23), use the <. statistic. 1 - . 1, Yr+3 - -1, and Y1+4 = -1 = 1, so that the sequence replicates itself every
Y1+2 -
fourth penod.
If you suspect a seasonal unit root, it is necessary to use an altemative procedure.
To keep the notation simple, suppose you ha ve quarterly observations on the {y,}
sequence and want to test for the presence of a seasonal unit root. To explain the T To devel.op the test, view (4.25) as a function of a ,, a 2· a 3, anda 4 and tak e a
methodology, note that the poi ynomial (1 - yL 4 ) can be factored, so that there are aylor senes ap roximation of A(L) around the point a = a = a = a = l.
1 2 3 4
four distinct characteristic roots: Although the deta1ls of the expansionare messy, first take the partial derivative:

(4.24) aA(L)laa 1 = a(l - a 1L)(l + a 2L)(l - a 3 iL)(l + a4 iL)laa

1
= -( 1 + a2L)(l - ail)(l + a 4 il)L
If y, has a seasonal unit root, y= l. Equation ( 4.24) is a bit restrictive in that it
only allows for a unit root atan annual frequency. Hylleberg et al. (1990) develop a Evaluating this derivative at the point a 1 = a2 = a 3 = a 4 = 1 yields
clever technique that allows you to test for unit roots at various frequencies; you
can test for a unit root (i.e., a root at a zero frequency), unit root at a semiannual -L(l + L)(l - il)(I + iL) = -L(I + L)(l + L2) =-L(l + L + L2'.ti.))
frequency, or seasonal unit root. To understand the procedure, suppose y, is gener-
Next, form
ated by

aA(L)!aa2 = a(I - a 1 L)(I + a 2L)(I - a 3 il)(I + a4il)!aa2

= (1 - a 1L)(l - a 3iL)(I + a4il)L
where A(l;) 1$ a fourth-order polynomial such that
Evaluating at the point 1 = a 2 = a 3 = a 4 = 1 yields (1 - L + L2 - L3)L. It should
(4.25) not ta e too long to convmce yourself that evaluating aA(L)!aa 3 and aA(L)/aa at
the pomt a 1 = a 2 = a 3 = a4 = 1 yields
4

Now, if a 1 = a 2 = a 3 = a 4 = 1, (4.25) is equivalent to setting y= 1 in (4.24).

Hence, if a 1 = a 2 = a 3 = a 4 = 1, there is a seasonal unit root. Consider sorne of the aA(L)laa3 = -(1 - L2)(1 + iL)iL
other possible cases:
and

CASE 1 aA(L)laa4 = ( 1 - L2)(1 - iL)iL

If a 1 = !, one homogeneous solution to (4.25) is y,= y,_ 1 • As such, the {y,} se- Since A(L) evaluated at ª1 = ª2 = a3 = a4 = 1 is (1 - L4), it is possible to approxi-
quence tends to repeat itself each and every period. This is the case of a nonsea- mate (4.25) by
sonal unit root.
[(l-L4)-L(I +L+L2+L3)(a1
2
- l)+(I-L+L2-L3)L(a2- l)

CASE 2 - (1 - L )(1 + il)il(a 3 - 1) + (1 - L2 )(1 - il)il(a 4 - l)]y, =E,

Define y, such that y, =(a, - 1) and note that (1 + iL)i = ¡- L and (1 _ iL)i = ¡+
If a 2 =
1, one homogeneous solution to (4.25) is y, + y,_ 1 = O. In this instance, the L; hence,
equence tends to replicate itself at 6-month intervals, so that there is a semiannual
unit root. For example, if y,= 1, it follows that Yr+i = -1, Yr+2 = + 1, Yr+J = -1, Yr+4 =
1, etc.
 Y1=0 Y2=0 Ys=r6=0
Intercept -2.88 -1.95 3.08
(1 - L4 )y, = y (1 +L.+ L
1
2 + L3 )y,_ 1 -y2 (1 - L + L2 -L3)y,_1 Intercept plus seasonal dummies -2.95 -2.94 6.57
. + (1 - L2)[()'3 - 1'4)i - ('{3 + )'4)L]y 1_1 + E 1 (4.26) Intercept plus seasonal dummies -3.53 -2.94 6.60
plus time trend
To purge the imaginary numbers from (4.26), define Ys and '{6 such. that 2'.J =
-y - iy and 2y =..;.y + iy . Hence, (y - y4 )i = y5 and y3 + '{4 = )'6. Subst1tutmg mto
6 5 4 6 5 3
(4.26) yields 4. EXAMPLES OF THE AUGMENTED DICKEY-FULLER
TEST

The last chapter reviewed the evidence reported by Nelson and Plosser (1982) sug-
gesting that macroeconomic variables are difference stationary rather than trend
Thus, to implement the procedure, use the following steps:
stationary. We are now in a position to consider their formal tests of the hypothesis.
For each series under study, Nelson and Plosser estimated the regression:
STEP 1: Form the following variables:

Y11- I = ( 1 + L + L2 + L1)y,_ I = Yr-1 + Yr-2 + _\'1-1 +y,. 4 p

2
hr-1 = (J - L + L - L1)Y1-I = Y1-I - Yr-2 + Yr-1 - Yr-4 L'i.yr =ªo + ª2t +'Y Y1-1 + L¡?.,-L'i.Y1-1+; +
i=2
E1

hr-I = ( 1 - L2)Yr-I = Yr-1 - Y1-J SO that Y \r-2 = Yr-2 - Yt-4

The chosen lag lengths are reported in the column labeled p in Table 4.2. The es-
STEP 2: Estimate the regression:
timated values a 0, a 2 , and y are reported in columns 3, 4, and 5, respectively.

You might want to modify the form of the equation by including an in- Table 4.2 Nelson and Plosser's Tests for Unit Roots
tercept, deterministic seasonal dummies, and a linear time trend. As in the
4
augmented form of the Dickey-Fuller test, lagged val u es of ( 1 - L )Y1-;
p
ªº ª2 r r+1
may also be included. Perform the appropriate diagnos ic checks t? ens re Real GNP 2 0.819 0.006 -0.175 0.825
that the residuals from the regression equation approx1mate a whtte-1101se (3.03) (3.03) (-2.99)
process. Nominal 2 1.06 0.006 -0.101 0.899
GNP >o•·..
(2.37) (2.34) (-2.32)
STEP 3: Form the t-statistic for the null hypothesis y 1 = O; the appropriate critica] Industrial 6 0.103 0.007 -0.165 0.835
"· · values are reported in Hylleberg et al. ( 1990). If you do not reject the hy- production (4.32) (2.44) (-2.53)
pothesis y, = O, conclude that a 1 = 1, so there is a nonseasonal unit root. Unemployment 4 0.513 --0.000 -0.294* 0.706
Next, form the t-test for the hypothesis y2 =O. If you do not reJeCt the null rate (2.81) (-0.23) (-3.55)
hypothesis, conclude that a 2 = 1 and there is a unit root with a semiannual
frequency. Finally, perform the F-test for the hypothesis Ys = y6 =O. If the Notes: l. p is the chosen lag length. Coefficients divided by their standard errors are in parentheses.
Thus. entries in parentheses represent the /-test for the null hypothesis that a coefficient is
calculated value is less than the critica! value reported 1n Hylleberg et al.
equal to zero. l)nder the null of nonstationary, it is necessary to use the Dickey-Fuller criti-
(1990), conclude that y5 and/or y is zero, so that there is a unit root with ca! values. At the 0.05 significance leve!, the critica! value for the t-statistic is -3.45.
6
an annual frequency. Be aware that the three null hypotheses are not alter- 2. An asterisk (*) denotes significance at the 0.05 leve!. For real and nominal GNP and indus-
natives; a series may have nonseasonal, semiannual, and annual umt roots. trial production, it is not possible to reject the null y= O at the 0.05 leve!. Hence, the unem-
ployment rate appears to be stationary.
At the 5% significance leve!, Hylleberg et al. ( 1990) report that the critica! val- 3. The expression y+ 1 is the estimate of the partía! autocorrelation between y, and y,_ 1 •
ues using 100 observations are:
/ Jesllng ;ar J renas una v111• /\ULJ•j

Recall that the traditional view of business cycles maintains that the GNP and In applied work, p, and P1 usually refer to national price indices in t relative to a
production Jevels are trend stationary rather than difference stationary. An adherent base year, so that e, refers toan index of the domestic currency price of foreign ex-
of this view must assert that y is different from zero; if y= O, the series has a unit change relative to a base year. For example, if the U.S. inflation rate is 10% while
root and is difference stationary. Given the sample sizes uscd by Nelson and the foreign inflation rate is 15%, the dallar price of foreign exchange should fall by
Plosser (1982), at the 0.05 leve!, the critica! value of the t-statistic for the null hy- approximately 5%. The presence of the term d, allows for short-run deviations from
pothesis y= O is -3.45. Thus, only if the estimated value of y is. more than 3.45 PPP.
standard deviations from zero, is it possible to reJeCt the hypothes1s that Y= O. As Because of its simplicity and intuitive appeal, PPP has been used extensively in
can be seen from inspection of Table 4.2. the estimated values of y for real GNP, theoretical models of exchange rate determination. However, as in the well-known
nominal GNP, and industrial production are not statistically different from zero. Dombusch (1976) "overshooting" model, real economic shocks, such as productiv-
Only the unemployment rate has an estimated value of y that is significantly differ- ity or demand shocks, can cause permanent deviations from PPP. For our purposes.
the theory of PPP serves as an excellent vehicle to illustrate many time-series test-
ent from zero at the 0.05 leve!.
ing procedures. One test of long-run PPP is to determine whether d, is stationary.
After ali, if the deviations from PPP are nonstationary (i.e., if the deviations are
Unit Roots and Purchasing-Power Parity
permanent in nature), we can reject the theory. Note that PPP does allow for persis-
Purchasing-power parity (PPP) is a simple relationship linking national price levels tent deviations; the autocorrelations of the {d,} sequence need not be zero. One
and exchange rates. In its simplest form, PPP asserts that the rate of currency depre- popular testing procedure is to define the "real" exchange rate in period t as
ciation is approximately equal to the difference betwecn the domestic and foreign
intlation rates. If p and p* denote the logarithms of the U.S. and foreign price levels
r, = e,+ P7 - p,
ande the logarithm of the dollar price of foreign exchange, PPP implies

e,= p, - P7 + d, Long-run PPP is said to hold if the { r1} sequence is stationary. For example, in
Enders (1988), I constructed real exchange rates for three major U .S. trading part-
ners: Germany, Canada, and Japan. The data were divided into two periods:
where d, represents the deviation from PPP in period t.
January 1960 to April 1971 (representing the fixed exchange rate period) and
January 1973 to November 1986 (representing the flexible exchange rate period).
Each nation's Wholesale Price Index (WPI) was multiplied by an index of the U.S.
dallar price of the foreign currency and then divided by the U.S. WPI. The log of
Figure 4.2 Real exchange rates. the constructed series is the {r,) sequence. Updated values of the real exchange rate
1.6 data used in the study are in the file REAL.PRN contained on the data disk. As an
exercise, you should use this data to verify the results reported below.
A critica! first step in any econometric analysis is to visually inspect the data.
1.4 The plots of the three real exchange rate series during the flexible exchange rate pe-
riod are shown in Figure 4.2. Each series seems to meander in a fashion characteris-
tic of a random walk process. Notice that there is Iittle visual evidence of explosive
1.2
behavior or a deterministic time trend. Consider Figure 4.3 that shows the autocor-
relation function of the Canadian real rate in levels, part (a), and first differences,
part (b). This autocorrelation pattem is typical of ali the series in the analysis. The
autocorrelation function shows little tendency to decay, whereas the autocorrela-
tions of the first differences display the classic pattem of a stationary series. In
0.8 graph (b), al! autocorrelations (with the possible exception of p 11 that equals 0.18)
are not statistically different from zero at the usual significance levels.
To formally test for the presence of a unit root in the real exchange rates, aug-
06
· 1973 1975 1977 1979 mented Dickey-Fuller tests of the form given by (4.19) were conducted. The re-
-Canada -Germany -Japan gression !:,.r, = a0 + yr,_ 1 + 2!:,.r,_ 1 + 3!:,.r,_2 + ... was estimated based on the follow-
ing considerations:
(Jan. 1973 = 1.00)
 i esttng ¡or J renas ana u1111 J<.uul>

Figure 4.3 ACF of Canada's real exchange rate. tistic for the null y= O is never more than l.59. The economic interpretation is
Levels. First dJfferences. that real productivity and/or demand shocks have had a pennanent influence on
1 1 1 1 1 1 1 1 1 1 1
real exchange rates.
0.4 2. As measured by the sample standard deviation (SD), real exchange rates were
0.8 - far more volatile in the 1973 to 1986 period than the 1960 to 1971 period.
0.2 Moreover, as measured by the standard error of the estimate (SEE), real ex-
0.6 change rate volatility is associated with unpredictability. The SEE during the
o flexible exchange rate period is severa! hundred times that of the fixed rate pe-
0.4 '--- riod. It seems reasonable to conclude that the change in the exchange rate
(0.2) regime (i.e., the end of Bretton-Woods) affected the volatility of the real ex-
0.2 - change rate.
(0.4) 3. Care must be taken to keep the appropriate null hypothesis in mind. Under the
o 1 2 3 4 5 6 7 8 9 10 11 12
null of a unit root, classical test procedures are inappropriate and we resort to the
1 2 3 4 5 6 7 8 9101112 statistics tabulated by Dickey and Fuller. However, classical test procedures
(a) (b) _(which assume stationary variables) are appropriate under the null that the real
rates are stationary. Thus, the following possibility arises. Suppose that the t-sta-
tistic in the Canadian case happened to be -2.16 instead of -1.42. Using the
1. The theory of PPP does not allow for a deterministic time trend or multiple unit Dickey-Fuller critica! values, you would not reject the null of a unit root; hence,
roots. Any such findings would refute the theory as posited. Although ali the se- you could conclude that PPP fails. However, under the null of stationarity
ries decline throughout the early l 980s and ali rise during the rnid to late l 980s, (where we can use classical procedures), y is more than two standard deviations
there is no a priori reason to expect a structural change. Pretesting the data using from zero and you would conclude PPP holds since the usual t-test becornes ap-
the Dickey-Pantula (1987) strategy showed no evidence of multiple unit roots. plicable.
Moreover, there was no reason to entertain the notion of trend stationarity; the This apparent dilemma commonly occurs when analyzing series with roots
expression a 2 t was not included in the estimating equation. close to unity in absolute value. Unit root tests do not have much power in dis-
criminating between characteristic roots close to unity and actual unit roots. The
2. In both time periods, F-tests and the SBC indicated that p3 through P 24 could be
dilemma is only apparent since the two null hypotheses are quite different. It is
set equal to zero. For Germany and Japan during the flexible rate period, 2 was
statistically different from zero; in the other four instances, p2 could be set equal perfectly consistent to maintain a null that PPP holds and not be able to reject a
to zero. In spite of these findings, with monthly data it is always important to null that PPP fails! Notice that this dilemma does not actually arise for any of
entertain the possibility of a lag length no shorter than 12 months. As such, tests the series reported in Table 4.3; for each, it is not possible to reject a null of
y= O at conventional significance levels.
were conducted using the short lags selected by the F-tests and SBC and using a
lag length of 12 months. 4. Looking at sorne of the diagnostic statistics, we see that ali the F-statistics indi-
cate that it is appropriate to exclude lags 2 (or 3) through 12 from the regression
For the Canadian case during the 1973 to 1986 period, the t-statistic for the null
equation. To reinforce the use of short lags, notice that the first-order correlation
hypothesis that y= O is -1.42 using no lags and -1.51 using ali 12 lags. Gi ven the
coefficient of the residuals (p) is low and the Durbin-Watson statistic close to 2.
critica! value of the 'tµ statistic, it is not possible to reject the null of a unit root in It is interesting that ali the point estimates of the characteristic roots indicate that
the Canadian/U.S. real exchange rate series. Hence, PPP fails for these two nations. real exchange rates are convergent. To obtain the characteristic roots, rewrite
In the 1960 to 1971 period, the calculated value of the t-statistic is -1.59; again, it
the estimated equations in the autoregressíve fonn r, = a0 + a 1 rr-1 or r, = a 0 +
is possible to conclude that PPP faíls. a 1 r,_1 + a 2 r,_ 2 . For the four AR( 1) models, the point estimates of the slope coef-
Table 4.3 reports the results of ali six estimations using the short lag lengths sug- ficients are ali less than unity. In the post-Bretton-Woods period (1973-1986),
gested by the F-tests and SBC. Notice the following properties of the estimated the point esti mates of the characteristic roots of Japan' s second-order process
models: are 0.931 and 0.319; for Germany, the roots are 0.964 and 0.256. However, this
l. For ali síx models, it is not possible to reject the null hypothesís that PPP fails. is precisely what we would expect if PPP fails; under the null of a unit root, we
As can be seen from the last column of Table 4.3, the absolute value of the t-sta- know that y is biased downward.
 °V': o
'T 5. PHILLIPS-PERRON TESTS
1

-
11
The distribution theory supporting the Dickey-Fuller tests assumes that the errors
are statistícally independent and have a constant variance. In using this methodol-
ogy' care-mu-st be taken to ensure that the error terrnsare uncorrelated and have
constant variance. Phillips and Perron ( 1988) developed a generalization of the
Dickey-Fuller procedure that allows for fairly mild assumptions conceming the
'° r--
00 -
distribution of the errors.
0 "'
N
To briefly explain the procedure, consider the following regression equations:

(4.27)

and

(4.28)

°o'
<n 00
::::: ó s 00
O"¡ r<l
00
ci ....:
00
ci 0i
- N
O N
o
ci ....: °' where T = number of observations and the disturbance term µ, is such that
1 1
Eµ,= O, but there is no requirement that the disturbance term is serially un-
correlated or homogeneous. Instead of the Dickey-Fuller assumptions of
independence and homogeneity, the Phillips-Perron test allows the dístur-
bances to be weakly dependent and heterogeneously distributed.
e
00
O': Phillips and Perron characterize the distributions and derive test statistics that
o
can be used to test hypotheses about the coefficients di' and ii; under the null hy-
pothesis that the data are generated by

r-- r-- y,=y,_¡ + µ,

°'
N 'T
N

9 9 The Phillips-Perron test statistics are modifications of the Dickey-Fuller t-statis-

tics that take into account the less restrictive nature of the error process. The ex-
pressions are extremely complex; to actually derive them would take us far beyond
00 o 'T
the scope of this book. However, many statistical time-series software packages
r-- r-- 00
now calculate these statistics, so that they are directly available. For the ambitious
°c'i °c'i °c'i reader, the formulas used to calculate these statistics are reported in the appendix to
this chapter. The most useful of the test statistics are as follows:

e
cr-, N
oo Z(tat): Used to test the hypothesis af = 1

1
ºoºo Z(tií 1 ): Used to test the hypothesis d 1 = 1
Z(tá 2 ): Used to test the hypothesis á2 =O
Z(<j> 3): Used to test the hypotheses d 1 = 1 and d2 =O

The critica! values for the Phillips-Perron statistics are precisely those given for
the Dickey-Fuller tests. For example, the critica! values for Z(taf) and Z(tá 1) are
e:
"o'. those given in the Dickey-Fuller tables under the headings 'tµ and 't,, respectively.
...., The critica! values of Z(<h) are given by the Dickey-Fuller <jl 3 statistic.
"'
 Do not be deceived by the apparent simplicity of Equations (4.27) and (4.28). In or
reality, it is far more general than the type of data-generating process allowable by
the Dickey-Fuller procedure. For example, suppose that the { µr) sequence is gener-
ated by the autoregressive process µr = (C(L)/B(L)]E,, where B(L) and C(L) are
Sr+90 - f r = Pt
polynomials in the lag operator. Given this fonn of the error process, we can write
Equation (4.27) in the forrn used in the Dickey-Fuller tests; that is, where Pr == per unit profit from speculation
E,p, =O

Thus, the efficient market hypothesis requires that for any time period t, the 90-
day forward rate (i.e., fr) be an unbiased estimator of the spot rate 90 days from t.
or
Suppose that a researcher collected weekly data of spot and forward exchange
rates. The data set would consist of the forward rates f r• Ír+?• f r+ 14, ••• and spot
(1 - af L)B(L)y, =a+ C( L)E,
rates s/! sr+?• sr+ 14 , • ••. By using these exchange rates, it is possible to construct the
sequence sr+ 9o - Ír = pf! s,+7 + 9o - Ír+? == Pr+?• sr+ 14+ 9o - Ír+t 4 == Pr+t4• .. .. Normalize
where a B(L) =a
the time period to 1 week, so that y 1 == pf! y 2 == Pr+?• y 3 == Pr+ 14, • •• and consider the
Thus, the Phillips-Perron procedure can be applied to mixed processes in the regression equation (where - is dropped for simplicity):
same way as the Dickey-Fuller tests.

Foreign Exchange Market Efficiency

The efficient market hypothesis asserts that ex ante expected profit must equal
Corbae and Ouliaris ( l 986) used Phillips-Perron tests to determine whether ( 1) ex- zero; hence, with quarterly data, it shou\d be the case that a 0 == a 1 == a 2 =O.
change rates follow a random walk and (2) the retum to forward exchange market However, the way that the data set was constructed means that the residuals will be
speculation contains a unit root. Denote the spot dollar price of foreign exchange on correlated. As Corbae and Ouliaris (1986) point out, suppose that there is relevant
<lay t as sr. An individual at t can also buy or sel! foreign exchange forward. A 90- exchange market "news" at date T. Agents will incorporate this news into a\l for-
day forward contract requires that on day t + 90, the individual take delivery (or ward contracts signed in periods subsequent to T. However, the realized retums for
make payment) of a specified amount of foreign exchange in retum for a specified ali preexisting contracts will be affected by the news. Since there are approximately
amount of dollars. Let f r denote the 90-day forward market price of foreign ex- 13 weeks in a 90-day period, we can expect the µr sequence to be an MA(l2)
change purchased on day t. On day t, suppose that an individual speculator buys process. Although ex ante expected returns may be zero, the ex post returns from
forward pounds at the price fr = $2.00/pound. Thus, in 90 days the individual is ob- speculation at t will be correlated with the returns from those engaging forward
ligated to provide $200,000 in return for Ll00,000. Of course, the agent may contracts at weeks t + 1 through t + 12.
choose to immediately sel! these pounds on the spot rnarket. lf on day t + 90, the Meese and Singleton (1982) assumed white-noise disturbances in using a
spot price happens to be sr+ 9o = $2.01/pound, the individual can sell the fl00,000 Dickey-Fuller test to study the returns from forward market speculation. One sur-
for $201,000; without transactíons costs taken into account, the individual earns
prising result was that the retum from forward speculation in the Swiss franc con-
a profit of $1000. In general, the profit on such a transaction will be sr+ 9 o - f,
tained a unit root. This finding contradicts the efficient market hypothesis since it
multiplied by the number of pounds transacted. (Note that profits will be negative if
implies the existence of a permanent component in the sequence of returns.
s,+90 <fr.) Of course, it is possible to speculate by selling forward pounds also. An
However, the assumption of white-noise disturbances is inappropriate if the {µ,}
individual selling 90-day forward pounds on day t will be able to buy them on the
sequence is an MA(12) process. Instead, Corbae and Ouliaris use the more appro-
spot market at sr+ 90 . Here, profits will be fr - Sr+ 9 o multiplied by the number of
priate Phillips-Perron procedure to analyze foreign exchange market efficiency;
pounds transacted. The efficient market hypothesis maintains that the expected
sorne of their results are contained in Table 4.4.
profit or loss from such speculative behavior must be zero. Let Er5r+ 9o denote the First, consider the test for the unit root hypothesis (i.e., a 1 == 1). All estimated
expectation of the spot rate for day t + 90 conditioned on the information available values of a 1 exceed 0.9; the first-order autocorrelation of the retums from specula-
on day t. Since we actually know fr on day t, the efficient market hypothesis for tion appears to be quite high. However, given the small standard errors, all esti-
forward exchange market speculation can be written as mated values are over four standard deviations from unity. At the 5% significance
leve!, the critica! value for a test of a 1 == 1, is -3.43. Note that this critica! value is
Er51+9o = Ír the Dickey-Fuller 't, statistic with 250 observations. Hence, as opposed to Meese
 m•vx

Table 4.4 Returns to Forward Speculation In practice, the choice of the most appropriate test can be difficult since you
./
.. ncver know the true data-generating process. A safe choice is to use both types of
/ ªº ª1 ªz unit roots tests. If they reinforce each other, you can ha ve confidence in the results.
Switzerland -D.117E-2 0.941 -0.11 !E-4 Sometimes, economic theory will be helpful in that it suggests the most appropriate
(0.106E-2) (0.159E-l) (0.834E-5) test. In the Corbae and Ouliaris example, excess returns should be positively corre-
Z(ta 0 ) = -1.28 Z(ta 1 ) = -4.06 Z(ta 2 ) = -1.07 latcd; hence, the Phillips-Perron test is a reasonable choice.
Canada -D.65 lE-3 0.907 O.l 16E-5
(0.409E-3) (0.191E-1) (0.298E-5)
Z(ta 0 ) = -1.73 Z(ta 1 ) = -5.45 Z(ta 2 ) = -1.42 6. STRUCTURALCHANGE
U.K. -D.779E-3 0.937 -D.132E-4
(0.903E-3) (0.163E-l) (0.720E-5) In performing unit root tests, special care must be taken if it is suspected that struc-
Z(ta 0) = -D.995 Z(la 1 ) = -4.69 Z(ta 2 ) = -1.50 tural change has occurred. When there are structural breaks, the various Dickey-
Notes: l. Standard errors are in parentheses. Fuller and Phillips-Perron test statistics are biased toward the nonrejection of a unit
2. Z(ta ) and Z(ta ) are the Phillips-Perron adjusted t-statistics for the hypothescs that a = O root. To explain, consider the situation in which there is a one-time change in the
0 2 0
and a 2 =O. respectively. Z(ta 1 ) is the Phillips-Perron adjusted r-statistic for the hypothesis mean of an otherwise stationary sequence. In the top graph (a) of Figure 4.4, the
that a,= l. {y,) sequen ce was constructed so as to be stationary around a mean of zero for t =
O, ... , SO and then to fluctuare around a mean of 6 for t = 51, ... , 1OO. The se-
and Singleton (1982), Corbae and Ouliaris are able to reject the null of a unit root quence was formed by drawing 100 normally and independently distributed values
in ali series examined. Thus, shocks to the retum from forward exchange market for the {E,) sequence. By setting Yo = O, the next 100 values in the sequence were
speculation do not ha ve permanent effects. generated using the formula:
A second necessary condition for the efficient market hypothesis to hold is that
the intercept term a0 equal zero. A nonzero intercept term suggests a predictable (4.29)
gap between the forward rate and spot rate in the future. If a0 =fa O, on average,
there are unexploited profit opportunities. lt may be that agents are risk-averse or where DL is a dummy variable such that DL =O for t = 1, ... , SO and DL = 3 for t =
profit-maximizing speculators are not fully utilizing ali available inforrnation in de- S 1, ... , 1OO. The subscript L is designed to indicare that the leve! of the dummy
terrnining their forward exchange positions. In absolute value, all the Z(td 0) statis- changes. At times, it will be convenient to refer to the value of the dummy variable
tics are less than the critica! value, so that Corbae and Ouliaris cannot reject the in period tas DL(t); in the example at hand, DL(SO) =O and DL(51) = 3.
null a 0 =O. In the same way, they are notable to reject the null hypothesis of no de- In practice, the structural change may not be as apparent as the break shown in
terrninistic time trend (i.e., that a 2 = 0). The calculated Z(tii2 ) statistics indicate that the figure. However, the large simulated break is useful for illustrating the problem
the estimated coefficients of the time trend are never more than l .SO standard errors of using a Dickey-Fuller test in such circumstances. The straight line shown in the
from zero. figure highlights the fact that the series appears to have a detenninistic trend. In
At this point, you might wonder whether it would be possible to perform the fact, the straight line is the best-fitting OLS equation:
same sort of analysis using an augmented Dickey-Fuller (ADF) test. After ali, Said
and Dickey (1984) showed that the ADF test can be used when the error process is
a moving average. The desirable feature of the Phillips-Perron test is that it allows
for a weaker set of assumptions conceming the error process. Also, Monte Cario
In the figure, you can see that the fitted value of a0 is negative and the fitted
studies find that the Phillips-Perron test has greater power to reject a false null hy-
value of a 2 is positive. The proper way to estímate (4.29) is to fit a simple AR(l)
pothesis of a unít root. However, there is a cost entailed with the use of weaker as- model and allow the intercept to change by including the dummy variable DL.
sumptions. Monte Cario studies have also shown that in the presence of negative However, suppose that we unsuspectingly fit the regression equation:
moving average terrns, the Phillips-Perron test tends to reject the null of a unit root
whether or not the actual data-generating process contains a negative unit root. It is
Yr = ªº + ªtYr-1 +e, (4.30)
preferable to use the ADF test when the true model contains negative moving aver-
age terms and the Phillips-Perron test when the true model contains positive mov- As you can infer from Figure 4.4, the estimated value of at is necessarily biased
ing average terms. toward unity. The reason for this upward bias is that the estimated value of a 1 cap-
 1
Figure 4.4 Two models of structural change.
Yr = Yo + ªot +LE¡
i=I

Thus, the misspecified equation (4.30) will tend to rnimic the trend line shown in
Figure 4.4 by biasing a 1 toward unity. This bias in a 1 means that the Dickey-Fuller
6 test is biased toward accepting the null hypothesis of a unit root, even though the
series is stationary within each of the subperiods.
Of course, a unit root process can exhibit a structural break also. The lower graph
4 (b) of Figure 4.4 simulates a random walk process with a structural change occur-
ring at t = 51. This second simulation used the same 100 realizations for the {E,)
sequence and set Yo= 2. The 100 realizations of the {y,} sequence were constructed
2 as

y,=y,_ 1 +E,+Dp

where D p(51) = 4 and ali other valu es of D P = O.

-2L- L- L- L- _... "------ --'
o 20 40 60 80 100 Here, the subscript P refers to the fact that there is a single pulse in the dummy
(a) variable. In a unit root process, a single pulse in the dummy will have a permanent
effect on the leve! of the {y,} sequence. In t = 51, the pulse in the dummy is equiva-
y,=Y,-1+E,+Dp lent to an E,+ 51 shock of four extra units. Hence, the one-time shock to D p(51) has a
8 permanent effect on the mean value of the sequence for t 51. In the figure, you
can see that the leve! of the process takes a discrete jump in t = 51, never exhibiting
any tendency to return to the prebreak leve!.
6 The bias in the Dickey-Fuller tests was confirmed in a Monte Cario experiment.
Perron (1989) generated 10,000 replications of a stationary process like that of
(4.29). Each replication was formed by drawing 100 normally and independently
4 distributed values for the {e,} sequence. For each of the 10,000 replicated series,
Perron used OLS to estimate a regression in the form of (4.30).8 As could be antici-
pated from our earlier discussion, he found that the estimated values of a 1 were bi-
2
ased toward unity. Moreover, the bias became more pronounced as the magnitude
of the break increased.
o
Testing for Structural Change
Retuming to the two graphs of Figure 4.4, we see that there may be instances in
-2 100
o 20 40 60 which the unaided eye cannot easily detect the difference between the alternative
(b) types of sequences. One econometric procedure to tests for unit roots in the pres-
ence of a structural break involves splitting the sample into two parts and using
tures the property that "low" values of y, (i.e., those fluctuating around zero) are Dickey-Fuller tests on each part. The problem with this procedure is that the de-
followed by other low values and "high" values (i.e., those fluctuating around a grees of freedom for each of the resulting regressions· are diminished. It is prefer-
mean of 6) are followed by other high values. For a formal demonstration, note that able to have a single test based on the full sample.
as a approaches unity, (4.30) approaches a random walk plus drift. We know Perron ( 1989) goes on to develop a formal procedure to test for unit roots in the
1 solution to the random walk plus drift model includes a deterrninistic trend, presence of a structural change at time period t = 't + 1. Consider the null hypothe-
that
the sis of a one-time jump in the leve! of a unit root process against the alternative of a
that is,
 _/
/

one-time change in the intercept of a trend stationary process. Formally, let the null STEP 1: Detrend the data by estimating the altemative hypothesis and calling the
and altemative hypotheses be residuals y,.
Henc , each val u e of y, is the residual from the regression y, = a 0 + a 2 t +
H 1 :y,=a 0 +yH +µ 1 Dp+t 1 (4.31)
µ2DL +y,. '
A 1 : y,= a0 + a 2 t + µ2DL +E, (432)

S1if> 2: Estimate the regression:

•re Dp represents a pulse dummy variable such that Dp = 1 if t = 't + 1 and
zero otherwise, and DL represents a leve! dummy variable such that DL = 1
if t > 't and zero otherwise.
Under the null hypothesis, {y,} is a unit root process with a one-time jump in the Under the null hypothesis of a unit root, the theoretical value of a1 is
unity. Perron ( 1989) shows that when the residuals are identically and
leve! of the sequence in period t = 't + l. Under the altemative hypothesis, {y,} is
independently distributed, the distribution of a 1 depends on the proportion
trend stationary with a one-time jump in the intercept. Figure 4.5 can help you to
of observations occurring prior to the break. Denote this proportion by:
visualize the two hypotheses. Simulating (4.31) by setting a0 = 1 and using 100 re-
"A= -r!T
alizations for the {E,} sequence, the erratic line in Figure 4.5 illustrates the time
path under the null hypothesis. You can see the one-time jump in the leve! of the where T= total number of observations.
process occurring in period 51. Thereafter, the {y,} sequence continues the original
random walk plus drift process. The altemative hypothesis posits that the {y,} se- .StlP 3: Perform diagnostic checks to determine if the residuals from Step 2 are se-
quence is stationary around the broken trend Iine. Up to t = 't, {y,} is stationary rially uncorrelated. If there is serial correlation, use the augmented form of
around a 0 + a 2 t and beginning 1: + l, y, is stationary around a0 + a 2t + µ2 . As ilius- the regression:
trated by the broken line, there is a one-time increase in the intercept of the trend if
µ2 >0. k
The econometric problem is to determine whether an observed series is best Yt = ª1Y1-I + I,!3;ó.Y1-i +El
modeled by (4.31) or (4.32). The implementation of Perron's (1989) technique is i=I
straightforward:
where y,= is the detrended series.
S&P 4: Calculare the t-statistic for the null hypothesis a 1 = l. This statistic can be
compared to the critica! values calculated by Perron. Perron generated
Figur:e 4.5
5000 series according to H 1 using values of "A ranging from O to 1 by
Alternative representations of structural change.
incre ments of 0.1. For each value of "A, he estimated the regressions y,
= a 1y,_ 1 + E, and calculated the sample distribution of a 1 • Naturally, the
critica! values are identical to the Dickey-Fuller statistics when f.. = O and
f.. = 1; in effect, there is no structural change unless O < f.. < 1. The maxi-
mum difference between the two statistics occurs when "A = 0.5. For f.. =
0.5, the critica! value of the t-statistic at the 5% leve! of significance is
-3.76 (which is !arger in absolute than the correspondíng Dickey-Fuller
statistic of -3.41). If you find a t-statistic greater than the critica! value
calculated by Perron, it is possible to reject the null hypothesis of a unit
roo t.
Of course, it is possible to incorporare Step 1 directly into Steps 2 or 3. To com-
bine Steps 1 and 3, simply estímate the equation:

Y1 =ªo +a,y,_1 +azt+µzDL + Ll3; Y1-i +Et

i=l
 /esttng1ur 1rena; ww """ l\uu•>

//'' The t-statistic for the null a 1 = 1 can then be compared to the appropriate critica! Perron's Test for Structural Changa
value calculated by Perron. In addition, the methodology is quite general in that it Perron ( 1989) used his analysis of structural change to challenge the findings of
can also allow for a one-time change in the drift or one-time change in both the Nelson and Plosser (1982). With the very same variables used, his results indicate
mean and drift For example, it is possible to test the null hypothesis of a permanent that most macroeconomic variables are not characterized by unit root processes.
change magnitude of the drift term versus the altemative of a change in the slope of Instead, the variables appear to be TS processes coupled with structural breaks.
the trend. Here, the null hypothesis is According to Perron (l989), the stock market crash of 1929 and dramatic oil price
increase of 1973 were exogenous shocks having permanent effects on the mean of
most macroeconomic variables. The crash induced a one-time fall in the mean.
Otherwise, macroeconomic variables appear to be trend stationary.
where DL = 1 if t > 't and zero otherwise. With this specification, the {y,} sequence Ali variables in Perron's study (except real wages, stock prices, and the station-
is generated by t:iy, = a0 +E, up to period r and t:iy, = (a0 + µ2 ) +E, thereafter. If µ1 > ary unemployment rate) appeared to have a trend with a constant slope and exhib-
O, the slope coefficient of the deterministic trend increases for t > 't. Similarly, a ited a major change in the level around 1929. In arder to entertain various hypothe-
slowdown in trend growth occurs if µ2 < O. ses concerning the effects of the stock market crash, consider the regression
The altemative hypothesis posits a trend stationary series with a change in the equation:
slope of the trend for t > 't:
k
y,= ªo +µ.,DL +µ.2Dp +a2t+a1Y1-1 + Lf3;f1Y1-i +E1
i=l
where DT = t - 't for t > 't and zero otherwise. For example, suppose that the break
occurs in period S 1 so that 't = 50. Thus, D,,.( 1) through D,,.(50) are all zero, so that
for the first 50 periods, {y,} evolves as y,= a 0 + a 2 t + E,. Beginning with period S 1, where Dp( 1930) = 1 and zero otherwise
D(S l)T = 1, D(S2)T = 2, ... , so that for t > 't, {y,} evolves as y 1 = a0 + (a2 + µ3)t + DL = l for ali t beginning in 1930 and zero otherwise
E,. Hence, DT changes the slope of the deterministic trend line. The slope of the Under the presumption of a one-time change in the mean of a unit root process,
trend is a2 for t:::; 't and a 2 + µ3 for t > 't. a 1 = 1, a 2 =O, and µ2 =O. Under the altemative hypothesis of a permanent one-time
To be even more general, it is possible to combine the two null hypotheses H 1 break in the trend stationary model, a 1 < 1 and µ1 =O. Perron' s {1989) results using
and H2 • A change in both the leve! and drift of a unit root process can be repre- real GNP, nominal GNP, and industrial production are reported in Table 4.5. Given
sented by the !ength of each series, the 1929 crash means that A. is 1/3 for both real and nomi-
nal GNP and egua! to 2/3 for industrial production. Lag lengths (i.e., the values of
k) were determined using t-tests on the coefficients ;· The value k was selected if
the t-statistic on k was greater than 1.60 in abso!ute value and the t-statistic on .
where Dp and DL = the pulse and leve! dummies defined above for i> k was less than 1.60.
The appropriate altemative for this case is First, consider the results for real GNP. When we examine the last column of the
table, it is clear that there is little support for the unit root hypothesis; the estimated
value of a 1 = 0.282 is significantly different from unity at the 1% level. Instead,
real GNP appears to have a deterministic trend (a2 is estimated to be over five stan-
dard deviations from zero). Al so note that the point estimate µ1 = -0.189 is signifi-
Again, the procedure entails estimating the regression A 2 or A 3 . Next using the
can ti y different from zero at conventional levels. Thus, the stock market crash is es-
residuals yl' estímate the regression: timated to have induced a permanent one-time decline in the intercept of real GNP.
These findings receive additional support since the estimated coefficients and
.Yr = ª1Y1-1 +e, their t-statistics are quite similar across the three equations. Ali values of a 1 are
about five standard deviations from unity, whereas the coefficients of the determin-
If the errors from this second regression equation do not appear to be white-
istic trends (a 2) are ali over five standard deviations from zero. Since ali estimated
noise, estímate the equation in the form of an augmented Dickey-Fuller test. The
t-statistic for the null hypothesis a 1 = 1 can be compared to the critica! values calcu- values of µ1 are significant at the 1% leve! and negative, the data seem to support
lated by Perron ( 1989). For A.= 0.5, Perron reports the critica! value of the t-statistic the contention that real macroeconomic variables are TS, except for a structural
at the 5% significance leve! to be -3.96 for H 2 and -4.24 for H3. break resulting from the stock market crash.
 .?"
/º"""6JV' •••"---··- •
>' Table 4.5 Retesting Nelson and Plosser's Data for Structural Change Dickey-Fuller tests yield

T A. k
ªº µ.¡ J1i ª2 ª1 y,= -Ü.0233y,_ 1 + E,, t-statistic for y= O: -0.98495
Real GNP 62 0.33 8 3.44 -0.189 -0.018 0.027 0.282 y,= 0.0661 - Ü.0566yl-I + E1, t-statistic for y= O: -1.70630
(5.07) (-4.28) (-0.30) (5.05) (-5.03) Yt = -Ü.0488 -Ü.1522y 1_¡ + 0.004t + E1, t-statistic for y= O: -2.73397
Nominal 62 0.33 8 5.69 -3.60 0.100 0.036 0.471
GNP (5.44) (-4.77) ( 1.09) (5.44) (-5.42) Diagnostic tests indicate that longer lags are not needed. Regardless of the p res-
ence of the constant or the trend, the {y,} sequence appears to be difference stat ion-
Industrial 111 0.66 8 0.120 -0.298 -0.095 0.032 0.322
production (4.37) (-4.56) (-.095) (5.42) (-5.47)
ary. Of course, the problem is that the structural break biases the data toward sug-
gesting a unit root.
Notes: 1.T = number of observations Now, with the Perron procedure, the first step is to estimate the model y f =a o +
A
A = proposition of observations occurring befare the structural change a2t + µ2DL + y,. The residuals from this equation are the detrended {Y,} sequence.
k = lag length
2. The appropriate 1-statistics are in parentheses. For a 0, µ,, µ,.anda,. the null is that the coef- The second step is to test for a unit root in the residuals by estimating yA 1 = a 1 yA t-1 +
ficient is equal to zero. For a 1 , the null hypothesis is a 1 = l. Note that al! estimated values of E1• The resulting regression is:
a 1 are significantly different from unity at the 1% leve!.
y,= 0.4843 y,_ 1 + E 1

In the third step, ali the diagnostic statistics indicate that {E,) approximates a
white-noise process. Finally, the t-statistic for a, = 1 is 5.396. Hence, we can reject
Tests with Simulated Data the null of a unit root and conclude that the simulated data are stationary around a
To further illustrate the procedure, 100 random numbers were drawn to represent breakpoint at t = 51.
the {E,} sequence. By setting y 0 == O, the next 100 values in the {y,} sequence were Sorne care must be used in using Perron's procedure since it assumes that the
drawn as date of the structural break is known. In your own work, if the date of the break is
uncertain, you should consult Perron and Vogelsang (1992). In fact, entire issue of
the July 1992 Journal of Business and Economic Statistics is devoted to break-
points and unit roots.

where D¿ == Ofort=l, ... ,50

D L = 1 for t == 51, ... , l 00 7. PROBLEMS IN TESTING FOR UNIT ROOTS
Thus, the simulation is identical to (4.29), except that the magnitude of the struc-
tural break is diminished. This simulated series is on the data file labeled There is a substantial literature concerning the appropriate use of the various
BREAK.PRN; you should try to reproduce the following results. If you were to plot Dickey-Fuller test statistics. The focus of this ongoing research concems the power
the data, you would see the same pattem as in Figure 4.4. However, if you did not of the test and presence of the deterministic regressors in the estimating equations.
plot the data or were otherwise unaware of the break, you might easily conclude Although many details are beyond the leve! of this text, it is important to be aware
that the {y,} sequence has a unit root. The ACF of the {y,} sequence suggests a unit of sorne of the difficulties entailed in testing for the presence of a trend (either de-
root process; for example, the first six autocorrelations are terministic or stochastic) in the data-generating process.

Lag: 1 2 3 4 5 6 Power
0.94 0.88 0.84 O.SI 0.77 0.72 Formally, the power of a test is equal to the probability of rejecting a false null hy-
pothesis (i.e., 1 minus the probability of a type II error). Monte Carlo simulations
and the ACF of the first differences is: have shown that the power of the various Dickey-Fuller and Phillips-Perron tests
is very low; unit root tests do not have the power to distinguish between a unit root
Lag: 1 2 3 4 5 6 and near unit root process. Thus, these tests will too often indicate that a series con-
-0.002 -0.201 -0.112 0.079 -0.010 -0.061 tains a unit root. Moreover, they have little power to distinguish between trend sta-
 ·onary and drifting processes. In finite samples, any trend stationary process can be Figure4.6
arbitrarily well approximated by a unit root process, and a unit root process can be
arbitrarily well approximated by a trend stationary process. These results should not Stationary and unit root processes.
be too surprising after examining Figure 4.6. The top graph (a) of the figure shows 3¡-- .,- -r-- .- -.-
a stationary process and unit root process. So as not to bias the results in any partic-
ular direction, the simulation uses the same 100 values of {E,} that were used in 2
Figure 4.4. Using these 100 realizations of {E,}, we constructed two sequences as:

y,= l.ly,_ 1 - 0.Jy,_2 + E1

z, = l.lz,_ 1 - 0.152 1_2 +E,
The {y,} sequence has a unit root; the roots of the {z,} sequence are 0.9405 and o
0.1595. Although {z,} is stationary, it can be called anear unit root process. If you
did not know the actual data-generating processes, it would be difficult to tell that -1
only {z,} is stationary.
Similarly, as illustrated in the lower graph (b) of Figure 4.6, it can be quite diffi-
cult to distinguish between a trend stationary and unit root plus drift process. Still -2
using the same l 00 values of {E1}, we can construct two other sequences as:

W, = 1 + 0.02! + E1 20 40 60 80 100
X1 = 0.02 + X 1_ 1 + E,/3 - Unit root process -Stationaryprocess

where x 0 = 1 (a)

Here, the trend and drift terms domínate the time paths of the two sequences.
Trend stationary and unit root processes.
Again, ·¡t is very difficult to distinguish between the sequences. This is especially
4r- --,-----,-----r-----r----
true since dividing each realization of E, by 3 acts to smooth out the {x,} sequence.
Just as it is difficult for the naked eye to perceive the differences in the sequences,
it is also difficult for the Dickey-Fuller and Phillips-Perron tests to select the cor-
rect specification. 3
It is easy to show that a trend stationary process can be made to mimic a unit root
process arbitrarily well. As discussed in Chapter 3, it is possible to write the ran-
dom walk plus noise model in the forrn:
2
y,=µ,+ T),
µ,=µ,_¡+E,

where T), and E, are both independent white-noise processes with variances of 0
and 02 , respectively. Suppose that we can observe the (y,} sequence, but
cannot directly observe the separate shocks affecting y,. If the variance of
E, is not zero, {y,} is the unit root process:

1
- Random walk plus drift - Trend stationary process
Yr = µo +LE¡ + T11 (4.33)
i l (b)
 On the other hand, if cr 2 = O, then ali values of {E,} are constant, so that: E, = the test. Reduced power means that the researcher may conclude that the process
E,_ 1 = ... = E:0 . To maintain the same notation as in previous chapters, define this ini- contains a unit root when, in fact, none is present. The second problem is that the
tial value of E0 as a 0 • lt follows that µ, = µ0 + a 0t, so {y,} is the trend stationary appropriate statistic (i.e., the 1, 1µ, and 1,) for testing y= O depends on which re-
process: gressors are included in the model. As you can see by examining the three Dickey-
Fuller tables, for a given significance leve!, the confidence intervals around y= O
(4.34) dramatically expand if a drift and time trend are included in the model. This is quite
different from the case in which {y,} is stationary. The distribution of the t-statistic
Thus, the difference between the difference stationary process of (4.33) and trend does not depend on the pre:;ence of the other regressors when stationary variables
process of (4.34) concems the variance of E1• Having observed the composite ef- are used.
fects of the two shocks-but not the individual components r¡, and E 1-we see that The point is that it is important to use a regression equation that mimics the ac-
there is no simple way to determine whether cr 2 is exactly equal to zero. This is par- tual data-generating process. If we inappropriately omit the intercept or time trend,
ticularly true if the data-generating process is such that cr is large relative to cr2 . In the power of the test cango to zero. 10 For example, if as in (4.35), the data-generat-
a finite sample, arbitrarily increasing cr will make it virtually impossible to distin- ing process includes a trend, omitting the term a 2t imparts an upward bias in the es-
guish between a TS and DS series.
It also follows that a trend stationary process can arbitrarily well approximate a
timated value of y. On the other hand, extra regressors increase the absolute value
of the critica! values so that you may fail to reject the null of a unit root. .,
unit root process. If the stochastic portian of the trend stationary process has suffi- To illustrate the problem, suppose that the time series {y,) is assumed to be gen-
cient variance, it will not be possible to distinguish between the unit root and trend erated by the random walk plus drift process:
!
stationary hypotheses. For example, the random walk plus drift model (a difference
astationary process) can be2 arbitrarily well represented by the model y, = a0 +
1y,_ 1 +E, by increasing cr and allowing a

these models can be approximated by (4.34).

1 to get sufficiently close to unity. Both

Does it matter that is often impossible to distinguish between borderline station-

a0 =1- Oanda 1 = l

where the inicial condition Yo is given and t = 1, 2, ... , T.

(4.36)

lf there is no drift, it is inappropriate to include the intercept term since the

l)
ary, trend stationary, and unit root processes? The realistic answer is that it depends power of the Dickey-Fuller test is reduced. When the drift is actually in the data-
on the question at hand. In borderline cases, the short-run forecasts from the alter- generating process, omitting a0 from the estimating equation also reduces the power
native models may have nearly identical forecasting performance. In fact, Monte
of the test in finite samples. How do you know whether to include a drift or time
Cario studies indicate that when the true data-generating process is stationary but trend in performing the tests? The key problem is that the tests for unit roots are
has a root close to unity, the one-step ahead forecasts from a differenced modelare
conditional on the presence of the deterministic regressors and tests for the pres-
usually superior to the forecasts from a stationary model. However, the Jong-run
ence of the deterministic regressors are conditional on the presence of a unit root.
forecasts of a model with a deterministic trend will be quite different from those of
Campbell and Perron (l991) report the following results conceming unit root tests:

-
the other models. 9

Determination of the Deterministic Regressors

J. When the estimated regression includes at least ali the deterministic elements in
the actual data-generating process, the distribution of y is nonnormal under the
null of a unit root. The distribution itself varies with the set of parameters in-
Unless the researcher knows the actual data-generating process, there is a question
cluded in the estimating equation.
conceming whether it is most appropriate to estimate (4.12), (4.13) or (4.14). It
2. lf the estimated regression includes deterministic regressors that are not in the
might seem reasonable to test the hypothesis y = O using the most general of the
models, that is, actual data-generating process, the power of the unit root test against a station-
ary altemative decreases as additional deterministic regressors are added.
p 3. If the estimated regression omits an important deterministic trending variable
6.yr =ao+'YYr-J +a2t+ I.,í3;6.Yt-i+I +E, (4.35) present in the true data-generating process, such as the expression a2t in (4.35),
;2 the power of the t-statistic test goes to zero as the sample size increases. If the
estimated regression omits a nontrending variable (i.e., the mean or a change in
After a!l, if the true process is a random walk process, this regression should find the mean), the t-statistic is consistent, but the finite sample power is adversely
that a0 = y= a 2 = O. One problem with this line of reasoning is that the presence of affected and decreases as the magnitude of the coefficient on the omitted com-
the additional estimated parameters reduces degrees of freedom and the power of ponent increases.

\
 . Estímating (4.13) or (4.14), we observe that the 'tµ, 't,, <jl 1 , <jl 2 , and <jl 3 statistics if the trend is significant, retest for the presence of a unit root (i.e., "( == 0)
have the asymptotic distributions tabulated by Dickey and Fuller ( 1979, 1981). using the standardized normal distribution. After all, if a trend is inappro-
The critica\ values of the various statistics depend on sample size. However, the priately included in the estimating equation, the limiting distribution of a2
sample variance of {y,} will be dominated by the presence of a trend or drift. is the standardized normal. If the nUll of a unit root is rejected, proceed no
We saw an example of this phenomenon in Figure 3.12 of Chapter 3. The time further; conclude that the {y,} sequence <loes not contain a unit root.
path of the random walk plus drift model in graph (b) is swamped by the pres- Otherwise, conclude that the {y,} sequence contains a unit root.
ence of the drift term. The fact that the stochastic trend is precisely the same as
in graph (a) has little effect on the overall appearance of the series. Although the Estimate (4.35) without the trend [i.e., estímate a model in the form of
proof is beyond the scope of this text, the 'tµ and 't, statistics converge to the (4.13)]. Test for the presence of a unit root using the 'tµ statistic. If the null
standardized normal. Specifically, is rejected, conclude that the model does not contain a unit root. If the null
hypothesis of a unit root is not rejected, test for the significance of the
constant (e.g., use the 'taµ statistic to test the significance of a0 given "( ==
0). Additional confirmation of this result can be obtained by testing the
t=I hypothesis a0 = y= O using the <P 1 statistic. If the drift is not significant, es-
',; ¡'
if a 0 7' O and a 2 = O timate an equation in the form of (4.12) and proceed to Step 4. If the drift
is significant, test for the presence of a unit root using the standardized
Only when both a0 and a 2 equal zero in the regression equation and data-gen-
erating process do the nonstandard Dickey-Fuller distributions dominate. If the data-
generating process is known to contain a trend or drift, the null hypothesis y= O Figure 4.7 A procedure to test for unit roots.
can be tested using the standardized normal distribution. Estimate L'.y, = a 0 + yy, -1 + ª2 t +I.{3¡L'.y, -i + Er

The direct implication of these four findings is that the researcher may fail to re- No STOP: Conclude
Is y= O? '
ject the null hypothesis of a unit root because of a misspecification conceming the no unit root.
deterministic part of the regression. Too few or too many regressors may cause a
¡Yes: Test for the presence No
failure of the test to reject the null of a unit root. Although we can never be sure of the trend.
that we are including the appropriate deterministic regressors in our econometric
model, there are sorne useful guidelines. Doldado, Jenkinson, and Sosvilla-Rivero Is a 2 =O No Is y= O Yes Conclude {y,) has
given using normal
(1990) suggest the following procedure to test for a unit root when the form of the a unit root.
y= O? distribution?
data-generating process is unknown. The following is a straightforward modifica-
tion of their method: ¡Yes

STEP 1: As shown in Figure 4.7, start with the least restrictive of the plausible Estimate No STOP: Conclude
L'.y, = ªº + yy, - 1 + I.f3J1y, - 1 +e,
i
models (which will generally include a trend and drift) and use the 't, sta- Is y= O?
no unit root.
tistic to test the null hypothesis y= O. Unit root tests ha ve low power to re-
Yes: Test for the presence No
ject the null hypothesis; hence, if the null hypothesis of a unit root is re-
of the drift.
jected, there is no need to proceed. Conclude that the (y,} sequence does
not contain a unit root. Is a 0 =O No lsy =O using Yes Conclude {y,) has
given ' normal
y= O? a unit root.
distribution?
STEP 2: If the null hypothesis is not rejected, it is necessary to determine whether
too many deterministic regressors were included in Step l above. 11 Test
for the significance of the trend term under the null of a unit root (e.g., use
i Yes

the 'tp, statistic to test the significance of a 2). You should try to gain addi- No Conclude
Estimate no unit root.
tional confirmation for this result by testing the hypothesis a2 = "( = O using L'.y,= YY1-1 +I./3¡L'.y, -1+e, Yes
the <j> 3 statistic. If the trend is not significant, proceed to Step 3. Otherwise, Is y= O? Conclude {y,} has
a unit root.
 normal. If the null hypothesis of a unit root is rejected, conclude that the where LGDP, == log(GDP,), so that MGDP, is the growth rate of real GDP, and
{y,} sequence does not contain a unit root. Otherwise, conclude that the standard errors are in parentheses.
{y,) sequence contains a unit root. The model is well estimated in that the residuals appear to be white-noise and ali
coefficients are of high quality. For our purposes, the interesting point is that the
STEP 4: Estimate (4.35) without the trend or drift, that is, esti mate a model in the Lllog(GDP,) series appears to be a stationary process. Integrating suggests that
form of (4.12). Use 1to test for the presence of a unit root. If the null hy- log(GDP,) has a stochastic and deterministic trend. The deterministic quarterly
pothesis of a unit root is rejected, conclude that the {y,} sequence does not growth rate of 0.007018--close to a 3% annual rate-appears to be quite reason-
contain a unit root. Otherwíse, conclude that the {y,} sequence contains a able. Now consider the three augmented Dickey-Fuller equations with t-statistics in
unit root. parentheses:

Remember, this procedure is not designed to be applied in a completely mechan-

ical fashion. Plotting the data is usually an important indicator of the presence of MGDP, == 0.79018 - 0.05409LGDP,_ 1 + 0.000348t
deterministic regressors. The data shown in Figure 4.1 could hardly be said to con- (2.56548) (-2.54309) (2.27941)
tain a detenninistic trend. Moreover, theoretical considerations might suggest the + 0.2496 lMGDP,_ 1 + 0.17273.:V.GDP,_ 2 (4.37)
appropriate regressors. The efficient market hypothesis is inconsistent with the (2.83349) (1.94841)
presence of a deterministic trend in an asset market price. However, the procedure RSS == 0.0089460783
is a sensible way to test for unit roots when the form of the data-generating process
is completely unknown. ALGOP, == 0.09600 - 0.0061lLGDP,_ 1 +0.23613.:V.GDP,_ 1
(2.05219) (-1.96196) (2.64113)
GDP and Unit Roots + 0.13535.:V.GDP,_ 2 (4.38)

Although the methodology outlined in Figure 4.7 can be very useful, it does have (1.52736)
its problems. Each step in the procedure involves a test that is conditioned on ali RSS = 0.0093334206

i the previous tests being correct; the significance leve! of each of the cascading tests
! is impossible to ascertain. MGDP, = 0.000279LGDP,_ 1 + 0.2633 lMGDP,_ 1 + 0.15443.:V.GDP,_ 2 (4.39)
l The procedure and its inherent dangers are nicely illustrated by trying to deter- (3.82135) (2.93959) (1.72964)
li mine if real gross domestic product (GDP) has a unit root. The data are contained in
the file entitled US.WKl on the data disk; it is a good idea to replicate the results
RSS = 0.0096582756
1
reported below. If we use quarterly data over the 1960:1to1991:4 period, the cor- From (4.37), the t-statistic for the null hypothesis y= O is -2.54309. Critica\ val-
relogram of the logarithm of real GDP exhibits slow decay. However, the first 12
j autocorrelations and partial autocorrelations of the logarithmic first difference are
ues with 125 usable observations are not reported in the Dickey-Fuller table. 12
However, with 100 observations, the critica] value of 1, at the 5% significance leve\
is -3.45; hence, it is not possible to reject the null hypothesis of a unit root given
the presence of the drift term and time trend.
\ ACF of the logarithmic first difference of real GDP:
Lag 1: 0.3093189 0.2316683 0.0572363 0.0556556 --0.0604932 0.0336679 The power of the test may have been reduced due to the presence of an unneces-
7: --0.0476200 --0.1453376 --0.0461222 0.0600729 0.0101171 -0.1695323 sary time trend and/or drift term. In Step 2, you test for the presence of the time
trend given the presence of a unit root. In (4.37), the t-statistic for the null hypothe-
PACF of the logarithmic first difference of real GDP: sis that a 2 =O is 2.27941. Do not Jet this large value fool you into thinking that a 2 is
Lag 1: 0.3093189 0.1503780 -0.0567524 0.0220048 --0.0876589 0.0696282 significantly different from zero. Remember, in the presence of a unit root, you
cannot use the critica] values of a Hable; instead, the appropriate critica] values are
7: --0.0507211 --0.1605942 0.0669240 0.1200468 -0.0353431 -0.2423071
given by the Dickey-Fuller 113, statistic. As you can see in Table 4.1, the critica!
value of 113, at the 5% significance leve] is 2.79; hence, it is reasonable to conclude
Despite the somewhat large partial correlation at lag 12, the Box-Jenkins procc-
that a 2 == O. The <j> 3 statistic to test the joint hypothesis a 2 == y== O reconfirms this re-
dure yields the ARIMA(O, 1, 2) model:
sult. If we view (4.37) as the unrestricted model and (4.39) as the restricted model,
there are two restrictions and 120 degrees of freedom in the unrestricted model; the
MGDP, == 0.007018 + (1+0.262169L + 0.197547L2)E,
<j> 3 statistic is
(0.001144) (0.088250) (0.082663)
 / ' = [(0.0096582756 -0.0089460783)/2] /(0.0089%0783/120) between real business cycles and the more traditional formulations, the nature of
- =4.7766 the trend may have important theoretical implications.
The usual t-statistics and F-statistics are not applicable to determine whether or
Since the critica! value of <!> 3 is 6.49, it is possible to conclude that the restriction nota sequence has a unit root. Dickey and Fuller (1979, 1981) provide the appro-
a2 = y = O is not binding. Thus, proceed to Step 3 where you estímate the model priate test statistics to determine whether a series contains a unit root, unit root plus
without the trend. In (4.38), the t-statistic for the null hypothesis y= O is -1.96196. drift, and/or unit root plus drift plus a time trend. The tests can also be modified to
Since the critica! value of the 'tµ statistic is -2.89 at the 5% significance leve!, the account for seasonal unit roots. If the residuals of a unit root process are heteroge-
null hypothesis of a unit root is not rejected at conventional significance levels. neous or weakly dependent, the altemative Phillips-Perron test can be used.
Again, the power of this test will have been reduced if the drift term does not be- Structural breaks will bias the Dickey-Fuller and Phillips-Perron tests toward
long in the model. To test for the presence of the drift, use the 'tm statistic. The cal- the nonrejection of a unit root. Perron (1989) shows how it is possible to incorpo-
culated t-statistic is 2.05219, whereas the critica! value at the 5% significance leve! rate a known structural change into the tests for unit roots. Caution needs to be ex-
is 2.54. The <!> 1 statistic also suggests that the drift term is zero. Comparing (4.38) ercised since it is always possible to argue that structural change has occurred; each
and (4.39), we obtain year has something different about it than the previous year. In an interesting exten-
sion, Perron and Vogelsang (1992) show how to test for a unit root when the pre-
qi 1 = (0.0096582756 - 0.0093334206)/(0.0093334206/121) cise date of the structural break is unknown.
= 4.21147365 Ali the aforementioned tests have very low power to distinguish between a unit
root and near unit root process. A trend stationary process, can be arbitrarily well
approximated by a unit. root process, and a unit root process can be arbitrarily well
Proceeding to Step 4 yields (4.39). The point is that the procedure has worked it-
approximated by a trend stationary process. Moreover, the testing procedure is con-
self into an uncomfortable comer. The problem is that the positive coefficient for y
founded by the presence of the deterministic regressors (i.e., the intercept and de-
(i.e., the estimated value of y= 0.000279 is almost four standard deviations from
terministic trend). Too many or too few regressors reduce the power of the tests.
zero) suggests an explosive process. In Step 3, it was probably unwise to conclude
An alternative is to take a Bayesian approach and avoid specific hypothesis test-
that the drift term is equal to zero. As you should verify in Exercise 4 at the end of
ing altogether. West and Harrison (1989) provide an accessible introduction to
this chapter, the simple Box-Jenkins ARIMA(O, 1, 2) model with an intercept of
Bayesian analysis in the context of regression analysis. Zellner (1988) discusses
0.007018 performs better than any of the altematives.
sorne of the philosophical underpinnings of the approach and Leamer (1986) pro-
vides a straightforward application to estimating the determinants of inflation. Sims
( 1988) is the standard reference for the Bayesian approach to unit roots.

SUMMARY ANO CONCLUSIONS

OUESTIONS ANO EXERCISES
In finite samples, the correlogram of a unit root process will decay slowly. As such,
a slowly decaying ACF can be indicative of a unit root or near unit root process.
1. The columns in the file labeled REAL.PRN contain the logarithm of the real
The issue is especially important since many economic time series appear to have a
ex- change rates for Canada, Japan, Germany, and the U.K. The four series
nonstationary component. When you encounter such a time series, do you detrend,
are called RCAN, RGER, RJAP, and RUK, respectively. As in Section 4,
do you first-difference, or do you do nothing since the series might be stationary?
each se- ries is constructed as r, =e,+ P1 - p,
Adherents of the Box-Jenkins methodology recommend differencing a nonsta-
tionary variable or variable with a near unit root. For very short-term forecasts, the where r = log of the real exchange rate
fonn of the trend is nonessential. Differencing also reveals the pattern of the other e = log of the dallar price of foreign exchange
autoregressive and moving average coefficients. However, as the forecast horizon p* = log of the foreign wholesale price index
expands, the precise form of the trend becomes increasingly important. Stationarity p = log of the U.S. wholesale price index
. implies the absence of a trend and long-run mean reversion. A deterministic trend
implies steady increases (or decreases) into the infinite future. Forecasts of a series Ali series run from February 1973 through December 1989, and each is ex-
with a stochastic trend converge to a steady leve!. As illustrated by the distinction pressed as an index number such that February 1973 = l .00.
.............

You should find that the data have the following properties:
D. If your software package can perform Phillips-Perron tests, reestimate part C
using the Phillips-Perron rather than Dickey-Fuller procedure. You should
Observa- Standard find that the t-statistics for y= O are
Series tions Mean Error Minimum Maximum
RCAN 203 0.93041911330 0.05685010789 0.83472000000 1.03930000000 Series No lags 12 lags Trend + 12 lags
RGER 203 1.07711822660 0.15732887872 0.64541000000 1.34009000000 RCAN -1.82209 -1.60022 -1.10882
RJAP 203 1.16689172414 0.13981473422 0.91620000000 1.50787000000 RJAP -1.82886 -2.03795 -2.19736
RUK 203 1.09026873892 0.14524762980 0.70991900000 1.38482000000 RGER -1.65117 -1.85319 -1.88371
A. For each sequence, find the ACF and PACF of (i) the leve! of the real ex- RUK -1.56654 -1.81530 -2.01424
change rate; (ii) the first difference of the real exchange rate; and (iii) the de-
trended real exchange rate. For example, for Canada you should find E. Why do you suppose the results from parts C and D are so similar?
F. Determine whether an intercept term belongs in the regression equations.
ACF:
Determine whether the time trend should be included in the equations.
1: 0.95109959 0.91691527 0.89743824 0.86897993 0.84708012 0.81911904 Determine whether the intercept and time trend belong in the equations.
7: 0.79706303 0.77888188 0.75410092 0.72946966 0.70020306 0.65782904
G. Use the Japanese data to show that you can reject the null hypothesis of two
unit roots.
ACF of the first difference:
1: --0.1562001 --0.1531103 0.0443029 --O.O 152957 0.1053500 -0.0740475 2. The second column in the file labeled BREAK.PRN contains the simulated data
7: --0.0475489 0.0597755 --0.0255490 0.0142241 0.1810469 --0.1151413 u sed in Section 6. You should find:

B. Explain why it is not possible to determine whether the seqence is stationary Observa- Standard
or nonstationary by the simple examination of the ACF and PACF. Series tions Mean Error Mínimum Maximum
C. Including a constant, use Dickey-Fuller and augmented Dickey-Fuller tests Yl 100 0.98802 0.99373 -0.78719 2.654697
(with 12 monthly lags) to test whether the series are unit root processes. You
should find that the t-statistics for y= O are A. Plot the data to see if you can recognize the effects of the structural break.
B. Verify the results reported in Section 6.
Series No lags 12 lags Trend + 12 lags
3. The third column in the file labeled BREAK.PRN contains another simulated
RCAN -1.81305 -1.50810 -0.85650
RJAP -1.81978 -2.30579 -2.61854 data set with a structural break at t = 51. You should find
RGER -1.64297 -2.10719 -2.09955
Observa- Standard
RUK -1.55877 -2.51668 -2.57493 Series tions Mean Error Mínimum Maximum
i. The last entry in the table means that y is more than 2.57 standard deviations Y2 100 2.21080 1.7816 -1.3413 5.1217
from zero. A student's t-table indicates that at the 95% significance leve!, the
critica! value is about 1.96 standard deviations. Why is it incorrect to conclude A. Plot the data. Compare your graph to those of Figures 4.4 and 4.5.
that the null hypothesis of a unit root can be rejected since the calculated 1-sta-
B. Obtain the ACF and PACF of the {Y2,} sequence and first difference of the
tistic is more than 1.96 standard deviations from zero? sequence. Do the data appear to be difference stationary?
11. For each entry reported in the table, what are the appropriate statistics to use (1:,
C. If as in (4.11), a Dickey-Fuller test is performed including a constant and
1:µ, or 1:,) in order to test the null hypothesis of a unit root? trend, you should obtain
 ··r1· .. ··-··•· · ......, .. · ..•....--· -····· · ·-·

Standard ENDNOTES
Coefficient Estimate Error t-Statistic Significance
s, serial
Constant 0.072445666 0.071447971 1.01396 l. Issues concerning the possibility of higher-order equations, longer lag length
0.31314869 correlation in the residuals, structural change, and the presence of detenninistic compo-
TREND -0.000101438 0.002120465 -0.04784 0.96194514 nents will be considered in due course.
Y2,_ 1 -0.022398360 0.034013944 -0.65851 0.51178974 2. The critica! values are reported in Table A at the end of this text.
3. Suppose that the estimated value of y is -1.9 (so that the estimate of a 1 is -0.9) with a
i. In what ways is this regression equation inadequate? standard error of 0.04. Since the estimated value of y is 2.5 standard errors from -2 [(2
- 1.9)/0.04 = 2.5], the Dickey-Fuller statistics indicate that we cannot reject the null
ii. What diagnostic checks would you want to perform? hy- pothesis y= -2 at the 95% significance leve!. Unless stated otherwise, the
discussion in the text assumes that a 1 is positive. Also note that if there is no prior
D. Estimate the. equation Y2, = a 0 + a 2 t + µ 2 DL and save tite residúals. You
information con- ceming the sign of a 1 , a two-tailed test can be conducted.
should obtain
4. Here we use the notation e,. rather than e,. to highlight that the residuals from such a re-
gression will generally not be white-noise.
Standard 5. For the same reason, it is also inappropriate to use one variable that is trend stationary
Coefficient Estima te Error t-Statistic Significancc and another that is difference stationary. In such instances, "time" can be included as a
Constant 0.4185991020 0.1752103414 2.38912 0.01882282 so-called explanatory variable or the variable in question can be detrended.
6. Tests using lagged changes in the {Ci.y,} sequence are called augmented Dickey-Fu!ler
DUMMY 2.8092054550 0.3097034669 9.07063 0.00000000
tests.
TREND 0.0076752509 0.0053644896 1.43075 0.15571516 7. In their simulations, Dickey and Fuller (1981) found that 90% of the calculated <j> 3 statís-
tícs were 5.47 or Jess and 95% were 6.49 or less when the actual data were generated ac-
E. Perform a pickey-Fuller test on the saved residuals. You should find y, =
cordíng to the null hypothesis.
0.965247ly,_ 1 +e,, where the standard error of a 1 =0.0372. Also perform the 8. Perron's Monte Cario study allows for a drift and deterministic trend. Nonetheless, the
appropriate diagnostíc tests on this regression to ensure that the residuals ap- value of a 1 is biased toward unity in the presence of the deterministic trend.
proximate white noise. You should conclude that the series is a unit root 9. Moreover, Evans and Savin (1981) find that for an AR(l) model, the li.niting distribu-
process with a one-time pulse at t = 51. tion of the autoregressive pararneter has a normal asymptotic distribution (for p < 1).
However, when the parameter is near l, the unit root distribution is a better finite sample
F. Retum to part D but now eliminate the insignificant time trend. How is your
approximation than the asymptotic correct distribution.
answer to part E affected?
10. Campbell and Perron (1991) report that omitting a variable that is growing at least as
fast as any other of the appropriately included regressors causes the power of the tests to
4. The sixth column in the file labeled US.WKl contains the real GDP data used in approach zero.
Section 7. The quarterly series runs from 1960:1to1991:4 and each entry is ex- 11. Using the most general model in Step l is meant to address the problem of omitting irn-
pressed in 1985 dollars. You should find that the properties of the series are such portant deterrnínistic regressors.
that 12. The sample period 1960:1to1991:4 contains 128 total observations. Three observations
are lost by creating the two lagged changes.
Series Name Observations Mean
GDP85 l28 3.220373E+ 12
APPENDIX: Phillips-Perron Test Statistics
A. Plot the logarithm of real GDP. Do the data suggest any particular form of
the trend? Suppose that we observe observations 1, 2, ... , T of the {y,} sequence and esti-
B. Use the Box-Jenkins methodology to verify that an ARIMA(O, 1, 2) model mate the regression equation:
performs better than an ARIMA(2, 1, 0) model.
C. Calculate the various Dickey-Fuller statistics reported in Section 7. Are
there any indications that might be inappropriate to accept the hypothesis In this appendix, we modify our notation slightly for those wishing to read the
a0 = O? work of Phillips and Perron. Fortunately, the changes are minor; simply replace ií 0
D. Repeat the procedure using the Phillips-Perron tests. with µ, ií 1 with a, and ií 2 with · Thus, suppose we have estimated the regression: