EViews Help - Unit Root Tests With A Breakpoint
EViews Help - Unit Root Tests With A Breakpoint
EViews Help - Unit Root Tests With A Breakpoint
User’s Guide : Advanced Univariate Analysis : Univariate Time Series Analysis : Unit Root Tests with a
Breakpoint
Unit Root Tests With A Breakpoint
Background
Break Variables
The Model
Innovational Outlier Tests
Additive Outlier Tests
Test Options
Lag Selection
Break Date Selection
Computing a Unit Root with Breakpoint Test
Examples
Real GNP
Employment
GNP Deflator
The use of unit root tests to distinguish between trend and difference stationary data has become
an essential tool in applied research. Accordingly, EViews offers a variety of standard unit root
tests, including augmented DickeyFuller (ADF), PhillipsPerron (PP), Elliot, Rothenberg, and Stock
(ERS), Ng and Perron (NP), and Kwiatkowski, Phillips, Schmidt, and Shin (KPSS) tests (“Unit Root
Testing”).
However, as Perron (1989) points out, structural change and unit roots are closely related, and
researchers should bear in mind that conventional unit root tests are biased toward a false unit
root null when the data are trend stationary with a structural break. This observation has spurred
development of a large literature outlining various unit root tests that remain valid in the
presence of a break (see Hansen, 2001 for an overview).
EViews offers support for several types of modified augmented DickeyFuller tests which allow for
levels and trends that differ across a single break date. You may compute unit root tests with a
single break where:
• The break can occur slowly or immediately.
• The break consists of a level shift, a trend break, or both a shift and break.
• The break date is known, or the break date is unknown and estimated from the data.
• The data are nontrending or trending.
Background
We begin with a brief discussion of the specifications underlining the testing methodology. As
always, our discussion is necessarily brief and we encourage you to consult the enclosed
references for additional detail.
Our discussion follows the basic framework outlined in Perron (1989), Vogelsang and Perron
(1998), Zivot and Andrews (1992), Banerjee et al. (1992) and others. For a useful overview of the
literature, see Perron (2006). Note that our notation differs slightly from the above sources.
Break Variables
http://www.eviews.com/help/helpintro.html#page/content%2FadvtimeserUnit_Root_Tests_with_a_Breakpoint.html%23 1/15
3/9/2017 EViews Help: Unit Root Tests with a Breakpoint
Before proceeding, it will be useful to define a few variables which allow us to characterize the
breaks. Let be an indicator function that takes the value 1 if the argument is true, and 0
otherwise. Then the following variables are defined in terms of a specified break date ,
• An intercept break variable
(38.27)
that takes the value 0 for all dates prior to the break, and 1 thereafter.
• A trend break variable
(38.28)
which takes the value 0 for all dates prior to the break, and is a break date rebased trend for
all subsequent dates.
• A onetime break dummy variable
(38.29)
which takes the value of 1 only on the break date and 0 otherwise.
Note that following EViews convention, we define the break date as the first date for the new
regime. This is in contrast to much of the literature which defines the break date as the last date
of the previous regime.
The Model
Following Perron (1989), we consider four basic models for data with a onetime break. For non
trending data, we have a model with (O) a onetime change in level; for trending data, we have
models with (A) a change in level, (B) a change in both level and trend, and (C) a change in
trend.
In addition, we consider two versions of the four models which differ in their treatment of the
break dynamics: the innovational outlier (IO) model assumes that the break occurs gradually,
with the breaks following the same dynamic path as the innovations, while the additive outlier
(AO) model assumes the breaks occur immediately. The tests considered here evaluate the null
hypothesis that the data follow a unit root process, possibly with a break, against a trend
stationary with break alternative.
Within this basic framework there are a variety of specifications for the null and alternative
hypotheses, depending on the assumptions one wishes to make about the break dynamics, trend
behavior, and whether the break date is known or determined endogenously.
As in Perron (1989), we consider two distinct approaches to modeling the break dynamics.
Innovational Outlier Tests
For the IO model, we consider the following general null hypothesis:
(38.30)
For our alternative hypothesis, we assume a trend stationary model with breaks in the intercept
and trend:
http://www.eviews.com/help/helpintro.html#page/content%2FadvtimeserUnit_Root_Tests_with_a_Breakpoint.html%23 2/15
3/9/2017 EViews Help: Unit Root Tests with a Breakpoint
(38.31)
with the breaks again following the innovation dynamics.
We may construct a general DickeyFuller test equation which nests the two hypotheses:
(38.32)
Within this general framework, we may specify different models for the null and alternative by
placing zero restrictions on one or more of the trend and break parameters , , . .
Following Perron (1989), Perron and Vogelsang (1992a, 1992b), and Vogelsang and Perron (1998),
we consider four distinct specifications for the DickeyFuller regression which correspond to
different assumptions for the trend and break behavior:
• Model 0: nontrending data with intercept break:
(38.33)
• Model 1: trending data with intercept break:
(38.34)
Setting the trend break coefficient to zero produces a test of a random walk with drift
against a trend stationary model with intercept break.
• Model 2: trending data with intercept and trend break:
(38.35)
The unrestricted DickeyFuller equation tests the random walk with drift against a trend
stationary with intercept and trend break alternative.
• Model 3: trending data with trend break:
(38.36)
Note that the test equation for Model 3 follows the methodology of Zivot and Andrews
(1992) and Banerjee et al. (1992) which does not nest the null and alternatives, as
is absent from the test equation; see Vogelsang and Perron (1998), p. 1077 for
discussion.
http://www.eviews.com/help/helpintro.html#page/content%2FadvtimeserUnit_Root_Tests_with_a_Breakpoint.html%23 3/15
3/9/2017 EViews Help: Unit Root Tests with a Breakpoint
You should bear in mind that whether one specifies a known break date or estimates the break
date from the data affects the allowable specifications for the null hypothesis.
If the break date is known as in Perron (1989), Models 0, 1, and 2 allow for breaks under the null
hypothesis. Model 3 does not allow for a break under the null.
If the break date is estimated, the test statistics considered here do not permit a breaking trend
under the null. Vogelsang and Perron (1998) offer a detailed discussion of this point, noting that
this undesirable restriction is required to obtain distributional results for the resulting Dickey
Fuller tstatistic. They offer practical advice for testing in the case where you wish to allow
under the null. See also Kim and Perron (2009) for more recent work that directly tackles this
issue.
Additive Outlier Tests
The general AO model null hypothesis is:
(38.37)
stationary and invertible ARMA error process, and is a drift parameter. Note that the full
impact of the break variables occurs immediately.
The alternative hypothesis is for a trend stationary model with possible breaks in the intercept
and trend:
(38.38)
Testing for a unit root in the AO framework is a twostep procedure where we first use the
intercept, trend, and breaking variables to detrend the series using OLS, and then use the
detrended series to test for a unit root using a modified DickeyFuller regression.
In the firststep of the AO test, we detrend the data using a model with appropriate trend and
break variables:
• Model 0: nontrending data with intercept break:
(38.39)
• Model 1: trending data with intercept break:
(38.40)
• Model 2: trending data with intercept and trend break:
(38.41)
• Model 3: trending data with trend break:
(38.42)
In the secondstep, let be the residuals obtained from the detrending equation. The resulting
DickeyFuller unit root test equation is given by,
• Models 0, 1, 2:
http://www.eviews.com/help/helpintro.html#page/content%2FadvtimeserUnit_Root_Tests_with_a_Breakpoint.html%23 4/15
3/9/2017 EViews Help: Unit Root Tests with a Breakpoint
(38.43)
• Model 3:
(38.44)
These are standard augmented DickeyFuller equations with the addition of break dummy
variables in Equation (38.44) to eliminate the asymptotic dependence of the test
statistic on the correlation structure of the errors and to ensure that the asymptotic distribution is
identical to that of the corresponding IO specification. See Perron and Vogelsang (1992b) for
discussion.
As with the IO tests, when we estimate the break date from the data, the distributional results
require that there be no trend break under the null hypothesis. See Vogelsang and Perron (1998)
and Kim and Perron (2009) for discussion.
Test Options
For a given test equation described above, you must choose a number of lags to include in the
test equation, and you must specify the candidate date at which to evaluate the break. EViews
offers a number of tools for you to use when making these choices.
Lag Selection
The theoretical properties of the test statistics requires that we choose the number of lag terms in
the DickeyFuller equations to be large enough to eliminate the effect of the correlation
structure of the errors on the asymptotic distribution of the statistic
• Fixed (with observationbased suggestion from Said and Dickey, 1984).
All of the remaining methods are data dependent, and require specification of a maximum lag
length . A different optimal lag length is obtained for each candidate break date.
• ttest.
Following Perron (1989), Perron and Vogelsang (1992a, 1992b), and Vogelsang and Perron
(1998), is chosen so that the coefficient on the last included dependent variable lag
difference is significant at a specified probability value, while the coefficients on the last
included lag difference in higherorder autoregressions up to are all insignificant at
the same level. The probability values for the tstatistics are computed using the t
distribution.
The ttest method requires the specification of a pvalue for use in evaluating significance.
The default pvalue of 0.10 may be changed by the user.
• Ftest.
Based on an approach of Said and Dickey (1984) (see also Perron and Vogelsang, 1992a,
1992b), the approach uses an Ftest of the joint significance of the lag coefficients for a
given against all higher lags up to . If any of the tests against higherorder lags
http://www.eviews.com/help/helpintro.html#page/content%2FadvtimeserUnit_Root_Tests_with_a_Breakpoint.html%23 5/15
3/9/2017 EViews Help: Unit Root Tests with a Breakpoint
are significant at a specified probability level, we set . If none of the test
statistics is significant, we lower by 1 and continue. We begin the procedure with
and continue until we achieve a rejection with , or until the
lower bound is evaluated without rejection and we set .
The Ftest method requires the specification of a pvalue for use in evaluating significance.
The default pvalue of 0.10 may be changed by the user.
• Information criterion
Following the approach of Hall (1994) and Ng and Perron (1995), is chosen to minimize
the specified information criterion amongst models with 0 to lags.
You may choose between the Akaike, Schwarz, HannanQuinn, Modified Akaike, Modified
Schwarz, Modified HannanQuinn. Note that the sample used for model selection excludes
data using full set of lag differences up to .
Break Date Selection
Perron (1989) specified an a priori fixed break date. Subsequent research (Zivot and Andrews,
1992; Banerjee et al., 1992; Vogelsang and Perron, 1998) has focused on endogenously
determining break dates from the data. EViews supports the following break date selection
methods:
• Minimize the DickeyFuller tstatistic .
Select the date providing the most evidence against the null hypothesis of a unit root and in
favor of the breaking trend alternative hypothesis.
).
Choose the date with the strongest evidence of a break. The alternative minimize and
maximize options are provided to allow for evaluation of onesided alternatives, and will
produce different critical values for the final DickeyFuller test statistic and tests with
greater power than the nondirectional alternatives.
• Userspecified break date.
Specify a fixed break date. This option allows you to carry out the original Perron (1989)
test.
For the automatic break selection methods, the following procedure is carried out. For each
possible break date, the optimal number of lags is chosen using the specified method, and the
test statistic of interest is computed. The procedure is repeated for each possible break date, and
the optimal break date is chosen from the candidate dates.
When the method is minimize , all possible break dates are considered. For the methods
http://www.eviews.com/help/helpintro.html#page/content%2FadvtimeserUnit_Root_Tests_with_a_Breakpoint.html%23 6/15
3/9/2017 EViews Help: Unit Root Tests with a Breakpoint
the break date.
Computing A Unit Root With Breakpoint Test
To compute a breakpoint unit root test, open a series window and select View/Breakpoint Unit
Root Test... to display the dialog:
The dialog is divided into six sections.
• The first section tells EViews whether you wish to compute the test using the raw data
(Level), or whether to test for higher order integration using differences (1st difference or
2nd difference) of the original data.
• The Trend specification section determines the trend components that are included in the
test. Using the Basic dropdown, you may choose between an Intercept only or an
Intercept and trend specification. If you include a trend in the specification you will be
prompted to indicate which deterministic components are breaking by choosing Intercept,
Intercept and trend, or Trend in the Breaking dropdown menu.
• The Lag length section describes the method for selecting lags for each of the
augmented DickeyFuller test specifications (“Lag Selection”). You may choose between
Akaike criterion (AIC), Schwarz criterion (BIC), HannanQuinn criterion (HQC),
Modified Akaike, Modified Schwarz, Modified HannanQuinn, tstatistic, Fstatistic,
and Fixed lag specifications. For all but the Fixed lag method, you must provide a Max. lag
to test; by default, EViews will suggest a maximum lag based on the number of
http://www.eviews.com/help/helpintro.html#page/content%2FadvtimeserUnit_Root_Tests_with_a_Breakpoint.html%23 7/15
3/9/2017 EViews Help: Unit Root Tests with a Breakpoint
observations in the series. For the test methods (tstatistic, Fstatistic), you must specify
a pvalue for the tests; for the Fixed lag method, you must specify the actual number of
use using the User lags edit field.
• The Break type section allows you to choose between the default Innovation outlier and
the Additive outlier specifications (“The Model”).
• The Breakpoint selection section specifies the method for determining the identity of the
breakpoint (“Break Date Selection”).
For a model with an intercept break, you may choose between minimizing the tstatistic for
in the ADF test (DickeyFuller mint), minimizing the tstatistic for the intercept break
coefficient (Intercept break mint), maximizing the tstatistic for the break coefficient
(Intercept break maxt), maximizing the absolute value of the tstatistic for the intercept
break coefficient (Intercept break maxabst), or providing a specific date (User
specified).
For models with a trend break, there will be corresponding entries for minimizing and
maximizing the tstatistic or absolute value of the tstatistic for the trend break coefficient.
For models with both an intercept and trend break you will be offered an additional choice
of using the Fstatistic for the break coefficients (Incpt.+trend break maxF) to select the
breakpoint.
You will be prompted for specify a trimming percentage when employing methods that involve
the tstatistic or Fstatistic of the break coefficients, EViews will remove from consideration
as the breakpoint this percentage of the observations from each endpoint.
For the Userspecified break choice you will be prompted to specify a single date.
• Lastly, the Additional output controls the output produced by the view. The checkbox
Display test and selection graphs controls whether to show only the test results with the
selected break, or to show the test results and graphs depicting the break selection criterion
results for each candidate break.
http://www.eviews.com/help/helpintro.html#page/content%2FadvtimeserUnit_Root_Tests_with_a_Breakpoint.html%23 8/15
3/9/2017 EViews Help: Unit Root Tests with a Breakpoint
If you provide a name in the Results matrix edit field, EViews will save the results from each
of the candidate augmented DickeyFuller tests in workfile. The first column contains the
observation identifier for the break; the second through fifth columns contain the
autoregressive coefficient, autoregressive coefficient standard error, number of
observations, number of variables, and number of selected lags in the DickeyFuller
regressions.
If appropriate, the remaining columns contain results for the breakpoint selection, with the
contents varying with the method chosen. When minimizing the DickeyFuller , the output
, the output contains the coefficient value, standard error, and the corresponding t
statistic; for the Fstatistic method, the output columns consist of the estimates of , the
standard error of , the estimates of , the standard error of , and the Fstatistic for
testing the significance of the two coefficients.
Examples
As examples, we replicate some of the results given in Perron (1997), using data originally
provided by Nelson and Plosser (1982). The dataset contains fourteen annual macroeconomic
series with values between 1860 and 1988. These data are provided in the workfile
“nelson_plosser.wf1”.
Real GNP
To begin, we replicate the results in the second row of Table 3 in Perron (1997), which tests for a
unit root in the log of real GNP using data between 1909 and 1970. We display the log of real
GDP, and set the workfile sample to dates from 1909 to 1970 with the commands
smpl 1909 1970
show log(rgnp)
To perform the unit root test with breakpoints, we click on View/Breakpoint Unit Root Test...
which brings up the test dialog. In this example Perron tests for the existence of a unit root of the
data in levels. The test assumes an innovation outlier break, with a trend specification given by
Model 2 (Equation (38.35), above); trending data with both intercept and trend break.
Perron selects a breakpoint by minimizing the DickeyFuller tstatistic, and selects a lag length
using the Ftest.
We can match these settings by clicking the Level and Innovation Outlier buttons, changing the
Basic Trend specification to Trend and Intercept and the Breaking Trend specification to
Intercept, selecting DickeyFuller mint as the Breakpoint selection, and changing the Lag
length Method to Fstatistic:
http://www.eviews.com/help/helpintro.html#page/content%2FadvtimeserUnit_Root_Tests_with_a_Breakpoint.html%23 9/15
3/9/2017 EViews Help: Unit Root Tests with a Breakpoint
Clicking OK produces the following results:
The top section of this output describes the test that was performed, with a description of the
underlying series, the trend and break specification, and the break type. The second section
displays the selected break date, which in this case is 1929. Recall that, unlike Perron, EViews
reports the break date for the start of the new regime instead of the last date before of the old
http://www.eviews.com/help/helpintro.html#page/content%2FadvtimeserUnit_Root_Tests_with_a_Breakpoint.html%23 10/15
3/9/2017 EViews Help: Unit Root Tests with a Breakpoint
regime, so the EViews reported date of 1929 matches Perron’s 1928 result. Lastly, we see that the
selected number of lags for corresponding test regression, selected on the basis of Fstatistic
selection is eight.
The lower section reports the Augmented DickeyFuller tstatistic for the unit root test, along with
Vogelsang’s asymptotic pvalues. Our test resulted in a statistic of 5.50, with a pvalue less than
0.01, leading us to reject the null hypothesis of a unit root.
EViews also provides a graph of the Augmented DickeyFuller statistics and AR coefficients at
each test date:
http://www.eviews.com/help/helpintro.html#page/content%2FadvtimeserUnit_Root_Tests_with_a_Breakpoint.html%23 11/15
3/9/2017 EViews Help: Unit Root Tests with a Breakpoint
Both graphs show a large dip in 1929, leaving little doubt as to which date should be selected as
the break point.
Employment
Our second example replicates row nine of Table 3 in Perron (2007). This example performs a
unit root test on the log of employment using data from 1890 to 1970. We again begin with
issuing commands to set the sample and display the log of employment:
smpl 1890 1970
show log(totalemp)
In this test, Perron again assumes an innovation outlier break, with a trend specification given by
Model 2 (Equation (38.35), above); trending data with intercept and trend break. However Perron
now selects the breakpoint corresponding to the minimum intercept break tstatistic, and selects
the laglength using the tstatistic method. We replicate these choices with the following dialog
settings:
http://www.eviews.com/help/helpintro.html#page/content%2FadvtimeserUnit_Root_Tests_with_a_Breakpoint.html%23 12/15
3/9/2017 EViews Help: Unit Root Tests with a Breakpoint
The first section of the results of this test are shown below:
Again, the top section of this output describes the test that was performed, notably the underlying
series, the trend and break specifications, and the break type. From the second section we can
see that again a date of 1929 was chosen as the most likely break date. The tstatistic based lag
selection selected seven lags for this test regression.
http://www.eviews.com/help/helpintro.html#page/content%2FadvtimeserUnit_Root_Tests_with_a_Breakpoint.html%23 13/15
3/9/2017 EViews Help: Unit Root Tests with a Breakpoint
The second section displays the test statistic and associated pvalue. The statistic value of 4.918
matches the value report by Perron, and the pvalue again means that we reject (at a 5%
significance level) the null hypothesis of a unit root.
GNP Deflator
Our final example replicates row 12 of Table 3 in Perron (1997), and performs a unit root test
with breaks on the log of the GNP deflator between 1889 and 1970. We set the workfile sample
and display the log of the GNP deflator by issuing the commands
smpl 1889 1970
show log(gnpdeflat)
We can mimic Perron’s results with the following dialog settings:
which yield results of:
http://www.eviews.com/help/helpintro.html#page/content%2FadvtimeserUnit_Root_Tests_with_a_Breakpoint.html%23 14/15
3/9/2017 EViews Help: Unit Root Tests with a Breakpoint
Here, 1920 was selected at the most likely break date, and the automatic lag selection routine
selected 9 lags.
The tstatistic of 3.869 matches that reported by Perron, and the corresponding pvalue of 0.27
indicates we cannot reject the hypothesis that the log of the GNP deflator has a unit root.
Last updated: Thu, 20 Jul 2017 13:44:40 PST
Back to top
http://www.eviews.com/help/helpintro.html#page/content%2FadvtimeserUnit_Root_Tests_with_a_Breakpoint.html%23 15/15