A PRACTITIONERS GUIDE TO

UNIT ROOT TESTING WITH TREND BREAKS

by

Lus Catela Nunes

Faculdade de Economia, Universidade Nova de Lisboa

April 22, 2005

1. Introduction

This quick practitioners guide provides some basic guidelines on conducting

ADF type tests for a unit root when there is a break in the trend function. It is based on
the results presented in Vogelsang and Perron (1998).

2. Trend Break Models

It is considered that a single break may have occurred in the trend function. The
date of the break is denoted by TB with 1 < TB < T , where T is the sample size. The
break is assumed to occur instantly.1 There are three possible cases given by,

(1) yt = + t + DU t + ut , (Model 1)

(2) yt = + t + DU t + DTt + ut , (Model 2)

(3) yt = + t + DTt + ut , (Model 3)

where,

0 if t TB ,
DU t =
1 if t TB + 1,

0 if t TB ,
DTt =
t TB if t TB + 1.

Model (1) allows for a break in the intercept. Model (2) allows for both a shift in the
intercept and slope. Model (3) allows for a smooth shift in the slope by requiring the
two segments of the broken trend to be joined.2

1
 The error ut is assumed to be an ARMA process.3

Under the alternative hypothesis, ut is stationary,4 so that yt is stationary

around a broken trend.

Under the null hypothesis of a unit root, ut has a unit autoregressive root, 5 so

that yt is I(1) and yt is a stationary process given by

yt = + Dt + vt , (Model 1)

yt = + Dt + DU t + vt , (Model 2)

yt = + DU t + vt , (Model 3)

where,

0 if t TB + 1,
Dt =
1 if t = TB + 1,

and the error vt is a stationary ARMA process.6

3. Unit Root Tests

Testing for a unit root in these models requires a two-step procedure. In the first
step, the series is detrended using one of the above equations: (1), (2) or (3). Denoting
by y%t the residuals from these regressions, in the second step, the unit root hypothesis is

tested using the t-statistic for testing = 1 in the regressions:7

y%t = y%t 1 + i =0 i Dt i + i =1 ci y%t i + t ,

k k
(Models 1,2)

y%t = y%t 1 + i =1 ci y%t i + t .

k
(Model 3)

The inclusion of the k lagged differences y%t i in all models, and of the k+1

dummy variables Dt i in Models 1 and 2, is necessary to ensure that the asymptotic

distribution of the t-statistics on are invariant to the correlation structure of the errors.

Next, two cases are considered. In the first case, the break date is known. In the
second case, the break date is unknown and must be estimated.

2
3.1. Unit Root Tests with a Known Break Date

The asymptotic critical values for the unit root tests with a known break date
appear in Table 1 and depend on the relative location of the break date in the sample
( TB /T). These critical values are valid if the assumed break date is correct or if there is
no break.8

Table 1. Asymptotic critical values for unit-root tests with a known break date
Model 1 Model 2 Model 3

TB /T 1% 2.5% 5% 10% 1% 2.5% 5% 10% 1% 2.5% 5% 10%

0.1 -4.30 -3.93 -3.68 -3.40 -4.38 -4.01 -3.75 -3.45 -4.15 -3.81 -3.52 -3.23
0.2 -4.39 -4.08 -3.77 -3.47 -4.65 -4.32 -3.99 -3.66 -4.34 -4.01 -3.72 -3.41
0.3 -4.39 -4.03 -3.76 -3.46 -4.78 -4.46 -4.17 -3.87 -4.41 -4.14 -3.85 -3.54
0.4 -4.34 -3.77 -3.72 -3.44 -4.81 -4.48 -4.22 -3.95 -4.48 -4.15 -3.91 -3.61
0.5 -4.32 -3.47 -3.76 -3.46 -4.90 -4.53 -4.34 -3.96 -4.49 -4.17 -3.93 -3.65
0.6 -4.45 -1.45 -3.76 -3.47 -4.88 -4.49 -4.24 -3.95 -4.50 -4.18 -3.94 -3.65
0.7 -4.42 -1.14 -3.80 -3.51 -4.75 -4.44 -4.18 -3.86 -4.49 -4.13 -3.89 -3.60
0.8 -4.33 -0.90 -3.75 -3.46 -4.70 -4.31 -4.04 -3.69 -4.41 -4.09 -3.83 -3.55
0.9 -4.27 -0.54 -3.69 -3.38 -4.41 -4.10 -3.80 -3.46 -4.29 -3.95 -3.72 -3.42
Note: Critical values for Models 1 and 2 are taken from Tables IV.B and VI.B respectively in Perron
(1989), and for Model 3 from Table II in Perron and Vogelsang (1993).

3
3.2. Unit Root Tests with an Unknown Break Date

When the break date TB is not known, Vogelsang and Perron (1998) suggest
choosing the break date that maximizes or minimizes a statistic that tests the
significance of one or more of the break parameters ( , ). Let t denote the t-statistic

for testing = 0 in Equations 1 or 2, t denote the t-statistic for testing = 0 in

Equations 2 or 3, and F , denote the F-statistic for testing = 0 and = 0 in Equation

2.

For Model 1, TB can be chosen using the maximum of | t | . If the direction of

the break is known to be positive ( > 0 ) then TB can be chosen using the maximum of

t . If the direction of the break is known to be negative ( < 0 ) then TB can be chosen

using the minimum of t .

For Model 2, TB can be chosen using the maximum of F , . TB can also be

chosen using the maximum of | t | . If the direction of the break is known to be positive

( > 0 ) then TB can be chosen using the maximum of t . If the direction of the break is

known to be negative ( < 0 ) then TB can be chosen using the minimum of t .

For Model 3, TB can be chosen using the maximum of | t | . If the direction of

the break is known to be positive ( > 0 ) then TB can be chosen using the maximum of

t . If the direction of the break is known to be negative ( < 0 ) then TB can be chosen

using the minimum of t .

When there is a break in the trend function, Vogelsang and Perron (1998) show
that using these procedures leads to the correct break date being chosen asymptotically.
Therefore, the asymptotic critical values presented in Table I are also valid.9

4. Lag Length Selection

To implement the above tests, the lag length k used in the second step
regressions must be chosen. For the case of a known break date, a model selection

4
criterion such as AIC or SC can be used.10 When the break date is estimated, this same
lag length selection procedure should be performed for the selected break date.

5
References

Perron, P. (1989). The great crash, the oil price shock, and the unit root hypothesis,
Econometrica 57(6), 1361-1401.

Perron, P. and T.J. Vogelsang (1993). The great crash, the oil price shock, and the unit
root hypothesis: Erratum, Econometrica 61(1), 248-249.

Vogelsang, T.J. and P. Perron (1998). Additional tests for a unit root allowing for a
break in the trend function at an unknown time, International Economic
Review, 39(4), 1073-1100.

1
This model is also called the additive outlier (AO) model. Another possible model is the innovational
outlier model where the break occurs more slowly over time. We only consider the AO model since it
has better asymptotic and finite sample properties, especially when large slope shifts are suspected.
Moreover, as shown in Vogelsang and Perron (1998), little is lost in terms of size and power if the AO
model is applied to IO data.
2
Models 1, 2 and 3 are denoted as Models A, C and B respectively in Perron (1989).
3
The error ut is specified to be an ARMA( p + 1, q ) process defined as (1 L ) A * ( L )ut = B ( L )et ,
where et is a white noise, and A * ( L ) is a pth order polynomial in L. It is assumed that A * ( L ) and
B ( L ) have all roots outside the unit circle.
4
The alternative hypothesis corresponds to | |< 1 , such that ut follows a stationary ARMA process.
5
The null hypothesis of a unit root corresponds to = 1 , such that ut follows a stationary ARMA
process: A * ( L ) ut = B ( L )et .
6
The null hypothesis of a unit root corresponds to = 1 , such that vt = ut follows a stationary ARMA
process: A * ( L )vt = B ( L )et .
7
The test statistic can also be computed as the t-statistic for = 0 in the following regressions:
y% t = y% t 1 + i Dt i + i =1 ci y% t i + t for Models 1 and 2; and in y% t = y% t 1 + i =1 ci y% t i + t
k k k

i =0

for Model 3.
8
If the true break date differs from the assumed break date then there will be some size distortions.
9
However, if the true data generating process has no break, but the above procedures are used to select
the break date, then the asymptotic distribution will be different. The corresponding critical values for
the case of no break are presented in Vogelsang and Perron (1998). From a practitioners point of view,
the existence of two alternative critical values for the same test raises a dilemma. There will be size
distortions in the following two situations: (1) if the critical values from Table I are used but the true
data generating process has no break; (2) if the critical values in Vogelsang and Perron (1998) for the
case of no break are used but there is a break. Because the critical values in Table I are less negative
than the critical values for the no break case in Vogelsang and Perron (1998), the true size will be larger
than the significance level in situation (1), and smaller in situation (2).
10 1/ 4
In EViews, by default, the maximum lag length considered equals | (12T / 100) |.