Applied Macro and

Financial Econometrics

Unit Root Test

Course Instructor:
Dr. Devasmita Jena
Unit Root Test: The origin
Consider:
--- (1)
If = 1, then eq(1) is RWM without a drift and it is NS
 = 1 : unit root problem, i.e., situation of NS
If 𝜌 = 1, eq(1) can be written as
• Solving characteristic equation (1- x)= 0, yields root 1, thus the name “unit root”
If <1
• Then is stationary
Types of unit root test:
• Dickey Fuller
• Augmented Dickey Fuller
• DF-GLS
• Phillips Perron
• KPSS
Dickey Fuller (DF) Test
Consider: --- (2)
If = 1, we have NS-condition, if <1, we have stationarity
We will also determine if (drift term) and (time trend term) are different from zero
Unit roots are usually tested by writing eq. (1) in 1st difference:
+ ( => + where = (
Testing for unit root boils down to testing for following set of hypotheses:
• H0: = 0
• H1: < 0
DF Test (Contd…)
Note:
• The test statistics is the usual t-statistic that is compared with the critical values from the DF tables(most
softwares use the Mckinnon critical values)
• H1 is ρ<0 or α3-1<0 or α3<1:
o this is stationarity condition of the AR(1) process: |α3|<1, i.e., coefficient of Y t-1 should be less than one
• H1 is one-sided:
• the rejection region is when the calculated value of t-statistic is smaller than the critical value where the values are to be taken with
their signs
o the rejection (acceptance) region is towards the left (right) of the critical value
o alternatively, accept the null of unit root if the calculated value of t-statistic is larger than the critical value
If H0 is rejected
• then, no unit root in the series => is stationary
If H0 is not rejected
• then ≃ I(d)
DF test needs to be repeated till you get I(0), accordingly equation (3) changes
We continue to perform the DF tests on higher order differencing of the series until one finds
that last differenced series to be stationary. Thus if {∆d Y } is I(0) then {Y } is I(d).
Augmented Dickey Fuller (ADF) Test
While applying the DF test the error term must be white noise while in models with higher
order autocorrelation terms for Yt (i.e.) in AR(p) models, the error term would be autocorrelated
The solution: use higher order differenced terms (that is lagged dependent variables) as
additional explanatory variables to approximate the autocorrelation

This test is called the ADF test

How many lags (k) to include?
 Information Criterion: SC, AIC, HQIC are used
 One could start with a reasonably large value of k and then systematically reduce the number of lagged terms
• Usually when one specifies the lag value the statistical software tends to report the AIC/ SC/ HQIC value for all the
lags up to the one specified by us and this could be used as a means to check for the correct lag length
 One could test for the joint significance of δi coefficient at each level of lag length using an F-test as well as
using the Durbin-Watson test for the error terms to check for no serial correlation.
• These two tests together will provide information on the correct specification and whether the error terms in the
model are white noise or not
ADF Test (Contd…)
Note:
• Too few lagged terms affects the size of the test (that is the null hypothesis of the series being I(1)
can be rejected too often)
• too many lagged terms result in not rejecting the null hypothesis often enough and also leads to
degree of freedom problem in a finite sample data
Test for integration:
• H0 : = 0; H1: < 0
• Accept H0 => presence of unit root
DF tables/Mckinnon critical values are used to compare the estimated and the tabulated values
DF test with MA Components:
• MA components in the ARIMA model are invertible
• Said and Dickey(1984) propose to approximate the ARMA structure by autoregressions of order k
where the lag length k grows with the sample size
• The ARIMA(p, 1, q) process can be adequately approximated by an ARIMA(n, 0, 0) process, where
n≤(T)1/3 ;T being the sample size
• Testing for a pure AR model is sufficient rather than for an ARIMA model
Seasonal Unit Roots
Suppose we have either a quarterly or a monthly time series
If this series has a unit root then it is represented as (respectively):
• Yt=Yt-4+εt
• Yt=Yt-12+εt
That is, the coefficient of the seasonal lag terms (t-4) or (t-12) (as may be appropriate) has a
value of 1.
The DF and ADF test has to be suitably modified to take this into account
A note about ADF: ADF incorporates RWM, RWMD, RWMD with a stochastic time trend
Phillips Perron Unit Root Test
Phillips and Perron (1988) developed a number of unit root as an advancement to ADF test
The PP unit root tests differ from the ADF tests mainly in how they deal with serial correlation
and heteroskedasticity in the errors
The ADF tests use a parametric autoregression to approximate the ARMA structure of the errors
in the test regression; PP test is non parametric version of ADF test
The modified test statistic:

The terms and are consistent estimates of variance parameters:

Phillips Perron Unit Root Test
The PP tests correct the ADF tests by the bias induced by the omitted autocorrelation
Under H0, the PP test statistics have the same asymptotic distributions as the ADF t-statistic and
normalized bias statistics
PP tests tend to be more powerful than the ADF tests
• Robust to general forms of heteroskedasticity in the error term εt
• No need to specify a lag length for the ADF test regression
But, they can severe size distortions (when autocorrelations of εt are negative) and they are
more sensitive to model misspecification (order of ARMA model)
Some Issues with Traditional Unit Root Tests
Power of tests is low if the process is stationary but with a root close to the non-stationary
boundary
Example, the tests are poor at distinguishing between =1 or =0.976, especially with small
sample sizes
• Suppose the true model is Yt = 0.976 Yt-1+ εt
• Then, H0: = 0 should be rejected
The ADF and PP unit root tests are known (from simulations) to suffer potentially severe
finite sample power and size problems
• Power – Both tests are known to have low power against the alternative hypothesis that the series is
stationary with a large autoregressive root. (See, DeJong, et al, J. of Econometrics, 1992.)
• Size – Both tests are known to have severe size distortion (in the direction of over-rejecting H0)
when the series has a large negative MA root (See, Schwert, JBES, 1989: MA = -0.8 =>size =
100%!)
DF-GLS Unit Root Test
Proposed by Elliott, Rothenberg and Stock (1996): an efficient test, modifying the DF test statistic using
a generalized least squares (GLS) rationale
Applies a GLS detrending step
Efficient in terms of small-sample size and improved power in the presence of an unknown mean or
trend
There are two possible alternative hypotheses: yt is stationary around a linear trend or Yt is stationary
with no linear time trend.
Under the first alternative hypothesis, the DF-GLS test is performed by first estimating the intercept and
trend via GLS.
 For a TS Yt, the GLS estimation is performed by generating the new variables, , , and where

, t=2,…T
DF-GLS Unit Root Test (Contd…)
An OLS regression is then estimated for the equation:

The estimated coefficients are then used to remove the trend from Y t

ADF test on the transformed variable, is performed, by fitting the OLS regression:

• H0: β = 0
To perform the DF-GLS test under the 2nd alternative hypothesis:
 proceed as above
 define α∗=1−(7/T)
 eliminate z from the GLS regression
 compute y∗ = yt −δ0
 fit the ADF regression by using the newly transformed variable
 perform the test - H0:β = 0
KPSS Unit Root Test
Developed by Kwiatkowski, Phillips, Schmidt and Shin(1992)
It can be used to test in the presence of both deterministic trend vs. stochastic trend
The KPSS test has the opposite null: the series being tested is stationarity
Consider:
is a RWM with initial value as , i.e., ; where ~ WN(0,σ 2), uncorrelated with ~WN(0,Ϙ2 )
Then,
H0(trend stationary): σ2=0
H0(Yt(level) stationary): σ2=0 & δ=0; against the alternative:
H1: σ2≠0, there is a RW in Yt
Under the assumptions of normality, i.i.d. for ut & εt test statistic null can be constructed with:
; where is the estimate of the long-run variance of the residuals and is the partial sum of the residuals:
Reject H0 when KPSS is large