RS – EC2 - Lecture 16

Lecture 16
Unit Root Tests

Autoregressive Unit Root

• A shock is usually used to describe an unexpected change in a
variable or in the value of the error terms at a particular time period.

• When we have a stationary system, effect of a shock will die out

gradually. But, when we have a non-stationary system, effect of a
shock is permanent.

• We have two types of non-stationarity. In an AR(1) model we have:

- Unit root: | 1 | = 1: homogeneous non-stationarity
- Explosive root: | 1 | > 1: explosive non-stationarity

• In the last case, a shock to the system become more influential as

time goes on. It can never be seen in real life. We will not consider
them.

1
RS – EC2 - Lecture 16

Autoregressive Unit Root

• Consider the AR(p) process:
 ( L ) yt     t where  ( L )  1  1 L1  L2 2  ....   p L p

As we discussed before, if one of the rj’s equals 1, Φ(1)=0, or

1   2  ....   p  1
• We say yt has a unit root. In this case, yt is non-stationary.

Example: AR(1): y t    1 y t 1   t  Unit root: 1 = 1.

H0 (yt non-stationarity): 1 = 1 (or, 1 – 1 = 0)
H1 (yt stationarity): 1 < 1 (or, 1 – 1 < 0)

• A t-test seems natural to test H0. But, the ergodic theorem and MDS
CLT do not apply: the t-statistic does not have the usual distributions.

Autoregressive Unit Root

• Now, let’s reparameterize the AR(1) process. Subtract yt-1 from yt:
yt  yt  yt 1    (1  1) yt 1   t
    0 yt 1   t
• Unit root test: H0: α0 = 1 – 1 = 0
H1: α0 < 0.

• Natural test for H0: t-test . We call this test the Dickey-Fuller (DF) test.
But, what is its distribution?

• Back to the general, AR(p) process: ( L) yt     t

We rewrite the process using the Dickey-Fuller reparameterization::

yt     0 yt 1  1yt 1   2 yt  2  ....   p 1yt ( p 1)   t

• Both AR(p) formulations are equivalent.

2
RS – EC2 - Lecture 16

Autoregressive Unit Root – Testing

• AR(p) lag Φ(L): ( L)  1  1 L1  L2  2  ....   p L p

• DF reparameterization:
(1  L )   0   1 ( L  L 2 )   2 ( L 2  L3 )  ....   p 1 ( L p 1  L p )

• Both parameterizations should be equal. Then, Φ(1)=-α0.

 unit root hypothesis can be stated as H0: α0=0.

Note: The model is stationary if α0< 0  natural H1: α0 < 0.

• Under H0: α0=0, the model is AR(p-1) stationary in Δyt. Then, if yt

has a (single) unit root, then Δyt is a stationary AR process.

• We have a linear regression framework. A t-test for H0 is the

Augmented Dickey-Fuller (ADF) test.

Autoregressive Unit Root – Testing: DF

• The Dickey-Fuller (DF) test is a special case of the ADF: No lags are
included in the regression. It is easier to derive. We gain intuition
from its derivation.

• From our previous example, we have:

 y t      1  y t 1   t     0 y t 1   t

• If α0 = 0, system has a unit root:: H 0 :  0  0

H1 :  0  0 (|  0 | 0)

ˆ  1
t  1 
• We can test H0 with a t-test:

SE ˆ

• There is another associated test with H0, the ρ-test:. (T-1)( ˆ −1).

3
RS – EC2 - Lecture 16

Review: Stochastic Calculus

• Kolmogorov Continuity Theorem
• If for all T > 0, there exist a, b, δ > 0 such that:
• E(|X(t1, ω) – X(t2, ω)|a) ≤ δ |t1 – t2|(1 + b)
• Then X(t, ω) can be considered as a continuous stochastic process.

– Brownian motion is a continuous stochastic process.

– Brownian motion (Wiener process): X(t, ω) is almost surely

continuous, has independent normal distributed (N(0,t-s)) increments
and X(t=0, ω) =0 (“a continuous random walk”).

Review: Stochastic Calculus – Wiener process

• Let the variable z(t) be almost surely continuous, with z(t=0)=0.
• Define (,v) as a normal distribution with mean  and variance v.
• The change in a small interval of time t is z

• Definition: The variable z(t) follows a Wiener process if

– z(0) = 0
– z = ε√t , where ε〜N(0,1)
– It has continuous paths.
– The values of z for any 2 different (non-overlapping) periods of
time are independent.
Notation: W(t), W(t, ω), B(t).

1
Example: WT (r )  (1   2   3  ...  [Tr ] ); r  [0,1]
T 8

4
RS – EC2 - Lecture 16

Review: Stochastic Process: Wiener process

• What is the distribution of the change in z over the next 2 time units?
The change over the next 2 units equals the sum of:
- The change over the next 1 unit (distributed as N(0,1)) plus
- The change over the following time unit --also distributed as N(0,1).
- The changes are independent.
- The sum of 2 normal distributions is also normally distributed.
Thus, the change over 2 time units is distributed as N(0,2).

• Properties of Wiener processes:

– Mean of z is 0
– Variance of z is t
– Standard deviation of z is √t
n
– Let N=T/t, then z (T )  z ( 0 )    t
i 1 i
9

Review: Stochastic Calculus – Wiener process

1 1
Example: WT (r )  (1   2   3  ...  [Tr ] )  S[Tr ] ; r  [0,1]
T T

• If T is large, WT(.) is a good approximation to W(r); r Є[0,1], defined:

W(r) = limT→∞ WT(r)  E[W(r) ] =0
 Var[W(r) ] =t

• Check Billingsley (1986) for the details behind the proof that WT(r)
converges as a function to a continuous function W(r).

• In a nutshell, we need
- εt satisfying some assumptions (stationarity, E[|εt|q <∞ for q>2, etc.)
- a FCLT (Functional CLT).
- a Continuous Mapping Theorem. (Similar to Slutzky’s theorem). 10

5
RS – EC2 - Lecture 16

Review: Stochastic Calculus – Wiener process

• Functional CLT (Donsker’s FCLT)
If εt satisfies some assumptions, then
WT(r) 
D
W(r),
where W(r) is a standard Brownian motion for r Є [0, 1].

Note: That is, sample statistics, like WT(r), do not converge to constants,
but to functions of Brownian motions.

• A CLT is a limit for one term of a sequence of partial sums {Sk},

Donsker’s FCLT is a limit for the entire sequence {Sk} instead of one
term.

11

Review: Stochastic Calculus – Wiener process

• Example: yt = yt-1 + εt (Case 1). Get distribution of (X’X/T2)-1 for yt.
T T t 1 T
T 2 ( y
t 1
t 1)
2
 T 2 
[ 
t 1 i 1
t i  y0 ]2  T 2 [S
t 1
t 1  y0 ]
2

T
 T 2 [(St 1
t 1)
2
 2y0 St 1  y02 ]

T T2
 St 1  1 S 
 2

t 1   T 
 T  2y0T 1/ 2  t 1 T 1  T 1 y02
t 1   T 

T t /T 2 T t /T
 1   1 
 2  
t 1 (t 1)T   T
S[Tr]  dr  2y0T 1/ 2

 
t 1 (t 1)T   T
S[Tr] dr  T 1 y02

1 1


 2 XT (r)2 dr  2y0T 1/ 2 XT (r) dr  T 1 y02
0
0
1


d

  2 W(r)2 dr, T  .
12
0

6
RS – EC2 - Lecture 16

Review: Stochastic Calculus – Ito’s Theorem

• The integral w.r.t a Brownian motion, given by Ito’s theorem (integral):
∫ f(t,ω) dB = ∑ f(tk,ω) ∆Bk where tk* є [tk,tk + 1) as tk + 1 – tk → 0.

As we increase the partitions of [0,T], the sum →p to the integral.

• But, this is a probability statement: We can find a sample path where

the sum can be arbitrarily far from the integral for arbitrarily large
partitions (small intervals of integration).

• You may recall that for a Rienman integral, the choice of tk* (at the
start or at the end of the partition) is not important. But, for Ito’s
integral, it is important (at the start of the partition).

• Ito’s Theorem result: ∫B(t,ω) dB(t,ω) = B2(t,ω)/2 – t/2. 13

Autoregressive Unit Root – Testing: Intuition

• We continue with yt = yt-1 + εt (Case 1). Using OLS, we estimate :
T T T

yy
t 1
t t 1  (y
t 1
t 1   y t 1 ) y t 1 y
t 1
t 1  y t 1
ˆ  T
 T
 1 T


t 1
2
y t 1 t 1
2
y t 1 y t 1
2
t 1

• This implies:
T T

y
t 1
t 1  y t 1  (y
t 1
t 1 / T )(  t / T )
T ( ˆ  1)  T T
 T

  (y
1
y 2t 1 t 1 / T )2
t 1
T t 1

• From the way we defined WT(.), we can see that yt/sqrt(T) converges
to a Brownian motion. Under H0, yt is a sum of white noise errors.

7
RS – EC2 - Lecture 16

Autoregressive Unit Root – Testing: Intuition

• Intuition for distribution under H0:
- Think of yt as a sum of white noise errors.
- Think of εt as dW(t).
Then, using Billingley (1986), we guess that T( ˆ -1) converges to
1

 W (t )dW (t )
T (ˆ  1) 
d 0
1

 W (t ) dt
2

• We think of εt as dW(t). Then, Σk=0 to t εk, which corresponds to

∫0to(t/T) dW(s)=W( s/T) (for W(0)=0). Using Ito’s integral, we have

1 W (1) 2  1
T (ˆ  1) 

d

21
 W (t ) dt
2

Autoregressive Unit Root – Testing: Intuition

• d 1 W (1) 2  1
T ( ˆ  1) 
2 1
W (t ) 2 dt
0

Note: W(1) is a N(0,1). Then, W(1)2 is just a χ2(1) RV.

• Contrary to the stable model the denominator of the expression

for the OLS estimator –i.e., (1/T)Σtxt2-- does not converge to a
constant a.s., but to a RV strongly correlated with the numerator.

• Then, the asymptotic distribution is not normal. It turns out that the
limiting distribution of the OLS estimator is highly skewed, with a
long tail to the left.

8
RS – EC2 - Lecture 16

Autoregressive Unit Root – Testing: Intuition

• DF distribution relative to a Normal. It is skewed, with a long tail to
the left.

Autoregressive Unit Root – Testing: DF

• Back to the AR(1) model. The t-test statistic for H0: α0=0 is given by
ˆ  1 ˆ  1
t 1  
 
SE ˆ T
s 2 ( y 2
)2
t 1
t2

• The test is a one-sided left tail test. If {yt} is stationary (i.e.,|φ| < 1)
then it can be shown

  d
 
T ˆ    N 0, 1   2 .
• This means that under H0, the asymptotic distribution of t=1 is
N(0,1). That is, under H0:
d
ˆ  N 1,0
which we know is not correct, since yt is not stationary and ergodic.

9
RS – EC2 - Lecture 16

Autoregressive Unit Root – Testing: DF

• Under H0, yt is not stationary and ergodic. The usual sample
moments do not converge to fixed constants. Using the results
discussed above, Phillips (1987) showed that the sample moments of
yt converge to random functions of Brownian motions. Under H0:
1

 W r dW r 

T  1  ˆ  1  d
0
1

 W r 
2
dr
0
1
d  W r dW r 
t 1  0
1/2
1

  W r  dr 
2

0 
where W(r) denotes a standard Brownian motion (Wiener process)
defined on the unit interval.

Autoregressive Unit Root – Testing: DF

• ˆ is not asymptotically normally distributed and tФ=1 is not
asymptotically standard normal.

• The limiting distribution of tФ=1 is the DF distribution, which does

not have a closed form representation. Then, quantiles of the
distribution must be numerically approximated or simulated.

• The distribution of the DF test is non-standard. It has been

tabulated under different scenarios.
1) with a constant: yt = μ + Φ yt-1 + εt.
2) with a constant and a trend: yt = μ + δ t +Φ yt-1 + εt.
3) no constant: yt = Φ yt-1 + εt.

• The tests with no constant are not used in practice.

10
RS – EC2 - Lecture 16

Autoregressive Unit Root – Testing: DF

• Critical values of the DFt test under different scenarios.

Autoregressive Unit Root – DF: Case 2

• Case 2. DF with a constant term in DGP: y t     y t 1   t
The hypotheses to be tested:
H 0 :   1,   0  Yt ~ I 1 without drift
H 1 :   1  Yt ~ I 0  with zero mean

This formulation is appropriate for non-trending economic and

financial series like interest rates, exchange rates and spreads.

• The test statistics tФ=1 and (T-1)(ˆ − 1) are computed from the
estimation of the AR(1) model with a constant.

11
RS – EC2 - Lecture 16

Autoregressive Unit Root – DF: Case 2

• Under H0 : Ф = 1, μ = 0, the asymptotic distributions of these test
statistics are influenced by the presence, but not the coefficient value,
of the constant in the test regression:
1

 W r dW r 


T  1ˆ  1 0 1
d

 W r  dr
 2

0
1

 W r dW r 

d
t 1  0
1/ 2
1  2 
  W r  dr 
 
0 
1 1
where W  r   W r    W r dr is a de - meaned Wiener process, i.e.,  W  r dr  0
0 0

• Inclusion of a constant pushes the tests’ distributions to the left.

Autoregressive Unit Root – DF: Case 3

• Case 3. With constant and trend term in the DGP.
The test regression is y t     t   y t 1   t

and includes a constant and deterministic time trend to capture the

deterministic trend under the alternative. The hypotheses to be tested:

H 0 :   1,   0  Yt ~ I 1 with drift

H 1 :   1  Yt ~ I 0  with determinis tic time trend

• This formulation is appropriate for trending time series like asset

prices or the levels of macroeconomic aggregates like real GDP. The
test statistics tФ=1 and (T −1) (ˆ − 1) are computed from the above
regression.

12
RS – EC2 - Lecture 16

Autoregressive Unit Root – DF: Case 3

• Again, under H0 : Ф = 1, δ = 0, the asymptotic distributions of both
test statistics are influenced by the presence of the constant and time
trend in the test regression. Now, we have:
1

 W r dW r 


T  1ˆ  1 0 1
d

 W r  dr
 2

0
1

 W r dW r 

d
t 1  0
1/ 2
1  2 
  W r  dr 
 
0 
 1  1
1
where W  r   W  r   12  r     s  W r dr is a de - meaned and de - trended
 2 0  2
Wiener process.

Autoregressive Unit Root – DF: Case 3

• Again, the inclusion of a constant and trend in the test regression
further shifts the distributions of tФ=1 and (T −1)(ˆ −1) to the left.

13
RS – EC2 - Lecture 16

Autoregressive Unit Root – Testing: DF

• Which version of the three main variations of the test should be
used is not a minor issue. The decision has implications for the size
and the power of the unit root test.

• For example, an incorrect exclusion of the time trend term leads to

bias in the coefficient estimate for Φ, leading to size distortions and
reductions in power.

• Since the normalized bias (T1)( ˆ− 1) has a well defined limiting
distribution that does not depend on nuisance parameters it can also
be used as a test statistic for the null hypothesis H0 : Ф = 1.

Autoregressive Unit Root – Testing: ADF

• Back to the general, AR(p) process. We can rewrite the equation as
the Dickey-Fuller reparameterization:
 y t     0 y t 1   1 y t 1   2  y t  2  ....   p 1 y t  ( p 1)   t

• The model is stationary if α0< 0 => natural H1: α0 < 0.

• Under H0: α0=0, the model is AR(p-1) stationary in Δyt. Then, if yt

has a (single) unit root, then Δyt is a stationary AR process.

• The t-test for H0 from OLS estimation is the Augmented Dickey-

Fuller (ADF) test.

• Similar situation as the DF test, we have a non-normal distribution.

14
RS – EC2 - Lecture 16

Autoregressive Unit Root – Testing: ADF

• The asymptotic distribution is:

The limit distributions DFα and DFt are non-normal. They are skewed
to the left, and have negative means.

• First result: α^0 converges to its true value (of zero) at rate T; rather
than the conventional rate of sqrt(T) => superconsistency.

• Second result: The t-statistic for α^0 converges to a non-normal limit

distribution, but does not depend on α.

Autoregressive Unit Root – Testing: ADF

• The ADF distribution has been extensively tabulated under the usual
scenarios: 1) with a constant; 2) with a constant and a trend; and 3)
no constant. This last scenario is seldom used in practice.

• Like in the DF case, which version of the three main versions of the
test should be used is not a minor issue. A wrong decision has
potential size and power implications.

• One-sided H1: the ADF test rejects H0 when ADF < c; where c is
the critical value from the ADF table.

Note: The SE(α^0)= s sqrt[Σyt-12], the usual (homoscedastic) SE. But,

we could be more general. Homoskedasticity is not required.

15
RS – EC2 - Lecture 16

Autoregressive Unit Root – Testing: ADF

• We described the test with an intercept. Another setting includes a
linear time trend:

 y t  1   2 t   0 y t 1   1 y t 1  ....   p 1 y t  ( p 1)   t

• Natural framework when the alternative hypothesis is that the series

is stationary about a linear time trend.

• If t is included, the test procedure is the same, but different critical

values are needed. The ADF test has a different distribution when t is
included.

Autoregressive Unit Root – DF: Example 1

• Monthly USD/GBP exchange rate, St, (1800-2013), T=2534.
- Case 1 (no constant in DGP):
Parameter Standard
Variable DF Estimate Error t Value Pr > |t|
x1 1 0.99934 0.00061935 1613.52 <.0001
(T-1)( -1)=2533*(1-.99934)=-1.67178
ˆ
Critical values at 5% level: 8.0 for T=500
8.1 for T=∞
• Cannot reject H0 => Take 1st differences (changes in St) to model
the series.

- With a constant, ˆ =0.99631. Similar conclusion (Critical values at

5% level: 14.0 for T=500 and14.1 for T=∞): Model changes in St.

16
RS – EC2 - Lecture 16

Autoregressive Unit Root – DF: Example 2

• Monthly US Stock Index (1800-2013), T=2534.
- No constant in DGP (unusual case, called Case 1): y t   y t 1   t
Parameter Standard
Variable DF Estimate Error t Value Pr > |t|
x1 1 1.00298 0.00088376 1134.90 <.0001
(T-1)( ˆ-1)=2533*(.00298)=7.5483 (positive, not very interesting)
Critical values at 5% level: 8.0 for T=500
8.1 for T=∞
• Cannot reject H0 => Take 1st differences (returns) to model the
series.

- With a constant, ˆ =1.00269. Same conclusion.

Autoregressive Unit Root – DF-GLS

• Elliott, Rothenberg and Stock (1992) (ERS) study point optimal
invariant tests (POI) for unit roots. An invariant test is a test invariant
to nuisance parameters.

• In the unit root case, we consider invariance to the parameters that

capture the stationary movements around the unit roots -i.e., the
parameters to AR(p) parameters.

• Consider: yt    t  u t , u t   u t 1   t .

• ERS show that the POI test for a unit root against ρ = ρ* is:

s21
MT 
s2*

17
RS – EC2 - Lecture 16

Autoregressive Unit Root – DF-GLS

• s21
MT 
s2*
where s2ρ is the variances residuals from the GLS estimation under
both scenarios for ρ, ρ = 1 and ρ = ρ*, respectively:

• The critical value for the test will depend on c where ρ* = 1 – c/T.

Note: When dynamics are introduced in the ut equation, ∆ ut lags, the

critical values have to be adjusted.

• In practice ρ* is unknown. ERS suggest different values for different

cases. Say, c=-13.5, for the case with a trend, gives a power of 50%.

Autoregressive Unit Root – DF-GLS

• It turns out that if we instead do the GLS-adjustment and then
perform the ADF-test (without allowing for a mean or trend) we get
approximately the POI-test. ERS call this test the DF-GLSt test.

• The critical values depend on T.

• Check ERS for critical values for other scenarios.

18
RS – EC2 - Lecture 16

Autoregressive Unit Root – Testing: ADF

• Important issue: lag p. :
- Check the specification of the lag length p. If p is too small, then the
remaining serial correlation in the errors will bias the test. If p is too
large, then the power of the test will suffer.

- Ng and Perron (1995) suggestion:

(1) Set an upper bound pmax for p.
(2) Estimate the ADF test regression with p = pmax.
If |tα(p)|> 1.6 set p = pmax and perform the ADF test.
Otherwise, reduce the lag length by one. Go back to (1)

- Schwert’s (1989) rule of tumb for determining pmax:

  T 1 / 4 
p max  12   
  100  

Autoregressive Unit Root – Testing: PP Test

• The Phillips-Perron (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. The PP tests
correct the DF tests by the bias induced by the omitted
autocorrelation.

• These modified statistics, denoted Zt and Z, are given by

.  ˆ 2 ˆ 2  T SE ˆ 
ˆ2 1     
Zt  t ˆ    0

ˆ 2 0 2  ˆ 2   ˆ 2 


1 T 2 SE ˆ 0  ˆ 2
Z   Tˆ 0 
2 ˆ 2
   ˆ 2 

19
RS – EC2 - Lecture 16

Autoregressive Unit Root – Testing: PP Test

• The terms ̂ 2 and ̂ are consistent estimates of the variance
parameters:

 E  
T T 1 T 
 2  lim T 1
T 
2
t  2  lim
T 
 E  T   2
t


t 1 t 1  t 1 

• Under H0:α0 = 0, the PP Zt and Zα0 statistics have the same

asymptotic distributions as the DF t-statistic and normalized bias
statistics.

• PP tests tend to be more powerful than the ADF tests. But, they can
severe size distortions (when autocorrelations of εt are negative) and
they are more sensitive to model misspecification (order of ARMA
model).

Autoregressive Unit Root – Testing: PP Test

• Advantage of the PP tests over 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.

20
RS – EC2 - Lecture 16

Autoregressive Unit Root – Testing: Criticisms

• The ADF and PP unit root tests are very popular. They have been,
however, widely criticized.

• Main criticism: Power of tests is low if the process is stationary but

with a root close to the non-stationary boundary.

• For example, the tests are poor at distinguishing between φ=1 or

φ=0.976, especially with small sample sizes.

• Suppose the true DGP is yt= 0.976 yt-1+ εt

=> H0:α0 = 0 should be rejected.

• One way to get around this is to use a stationarity test (like KPSS
test) as well as the unit root ADF or PP tests.

Autoregressive Unit Root – Testing: Criticisms

• The ADF and PP unit root tests are known (from simulations) to
suffer potentially severe finite sample power and size problems.

1. Power – Both tests are known to have low power against the
alternative hypothesis that the series is stationary (or TS) with a large
autoregressive root. (See, DeJong, et al, J. of Econometrics, 1992.)

2. 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%!)

21
RS – EC2 - Lecture 16

Autoregressive Unit Root – Testing: KPSS

• A different test is the KPSS (Kwiatkowski, Phillips, Schmidt and
Shin) Test (1992). It can be used to test whether we have a
deterministic trend vs. stochastic trend:
H 0 : Yt ~ I 0   level or trend  stationary
H 1 : Yt ~ I 1  difference stationary
• Setup
y t     t  rt  u t
rt  rt  1   t

where εt ~ WN(0,σ2), uncorrelated with ut ~WN. Then,

H0(trend stationary): σ2=0
H0(yt (level) stationary): σ2=0 & δ=0.

Under H1: σ2≠0, there is a RW in yt .

Autoregressive Unit Root – Testing: KPSS

• Under some assumptions (normality, i.i.d. for ut & εt), a one-sided
LM test of the null that there is no random walk (εt=0, for all t) can be
constructed with:
T
St
KPSS  T 2

t 1 su 2

where su2 is the variance of ut, (“long run” variance) estimated as

su2(l)

where w(s,l) is a kernel function, for example, the Bartlett kernel. We

also need to specify the number of lags, which should grow with T.

• Under H0, uෝt can be estimated by OLS.

22
RS – EC2 - Lecture 16

Autoregressive Unit Root – Testing: KPSS

• Easy to construct. Steps:
1. Regress yt on a constant and time trend. Get OLSt residuals, u^.
2. Calculate the partial sum of the residuals: S t   uˆ i
i 1
3: Compute the KPSS test statistic
T
St
KPSS  T 2

t 1 su 2

where su2 is the estimate of the long-run variance of the residuals.

4. Reject H0 when KPSS is large (the series wander from its mean).

• Asymptotic distribution of the test statistic is non-standard –it can

be derived using Brownian motions, appealing to FCLT and CMT.

Autoregressive Unit Root – Testing: KPSS

• KPSS converges to three different distribution, depending on
whether the model is trend-stationary (δ≠0),, level-stationary (δ=0),
or zero-mean stationary (δ=0, μ=0).

• For example, if a constant is included (δ=0) KPSS converges to

1
d
KPSS  [W r   W (1)] dr

0

Note: V = W(r) − rW(1) is called a standard Brownian bridge. It

satisfies V(0)=V(1)=0.

• It is a very powerful unit root test, but if there is a volatility shift it

cannot catch this type non-stationarity.

23
RS – EC2 - Lecture 16

Autoregressive Unit Root – Structural Breaks

• A stationary time-series may look like non-stationary when there are
structural breaks in the intercept or trend.

• The unit root tests lead to false non-rejection of the null when we do
not consider the structural breaks. A low power problem.

• A single known breakpoint was studied by Perron (Econometrica,

1989). Perron (1997) extended it to a case of unknown breakpoint.

• Perron considers the null and alternative hypotheses

H0: yt = a0 + yt-1 + µ1DP + εt (yt ~ ST with a jump)
H1: yt = a0 + a2t + µ2DL + εt (yt ~ TS with a jump)

Pulse break: DP = 1 if t = TB + 1 and zero otherwise,

Level break: DL = 0 for t = 1, . . . , TB and one otherwise.

Unit Root – Single Structural Break: Perron

•. Power of ADF tests: Rejection frequencies of ADF–tests

• Observations:
- ADF tests are biased toward non-rejection of the non-stationary H0.
- Rejection frequency is inversely related to the magnitude of the shift.

• Perron estimated values of the AR coefficient in the DF regression.

They were biased toward unity and that this bias increased as the
magnitude of the break increased.

24
RS – EC2 - Lecture 16

Unit Root – Single Structural Break: Perron

• Perron’s suggestion: Running the following OLS regression:
p
y t  a 0  a 1 y t 1  a 2 t   2 D L    i  y t  i   t
i 1
H0: a1=1;  use t–ratio, DF unit root test.

• Perron shows that the asymptotic distribution of the t-statistic

depends on the location of the structural break,  = TB/T.

• Perron (1989) derives critical values for different cases. For example:
H0: yt = a0 + yt-1 + µ1DP + γ1 DL + εt (yt ~ ST with a jump & break)
H1: yt = a0 + a2t + µ2DL + γ2 DL + εt (yt ~ TS with a jump & break)

• Main problem with this test procedure: structural breaks are not
known, they need to be estimated from data.

Unit Root – Single Structural Break: ZA

• Main problem with this test: Structural breaks are not known, they
need to be estimated. Many papers dealing with endogenous structural
breaks: Zivot and Andrews (ZA, 1992), Lumsdaine and Papell (1998),
Lee and Strazicich (2003).

• ZA’s test is a sequential ADF test, using a different dummy variable

for each possible break date. The break date is selected where the t-
statistic from the ADF test is a minimum (most negative) –break date
is chosen where the evidence is least favorable for the unit root null.

• ZA’s critical values are different from Perron’s (1989). In general,

ZA provide more evidence for unit roots than under Perron’s.

25
RS – EC2 - Lecture 16

Unit Root – Multiple Structural Breaks

• Lumsdaine and Papell (1998) and Lee and Strazicich (2003) allow for
multiple breaks in their tests.

• Lumsdaine and Papell extend ZA, by allowing two structural breaks

under the alternative hypothesis of the unit root test and additionally
allow for breaks in level and trend.

• The derivation of critical values on ZA and Lumsdaine and Papell

(1998) assumes no breaks under the null hypothesis. This assumption
may lead to conclude incorrectly (spuriously) reject H0 (unit root) when,
in fact, the series is difference-stationary with breaks.

Unit Root – Multiple Structural Breaks

• The derivation of critical values on ZA and Lumsdaine and Papell
(1998) assumes no breaks under the null hypothesis. This assumption
may lead to conclude incorrectly (spuriously) reject H0 (unit root) when,
in fact, the series is difference-stationary with breaks .

• To deal with this issue, Lee and Strazicich (2003) propose a LM unit-
root test, incorporating structural breaks under H0 (& H1), with
DGPs (augmenting with p first-difference AR lags works well):

H0: yt = a0 + yt-1 + µ1DP,1 + µ1DP,2 + γ1 DL,1 + γ2 DL,2 + εt

H1: yt = a0 + (1-a1) yt-1 + a2t + µ1DP,1 + µ1DP,2 + γ1 DL,1 + γ2 DL,2 + εt
• In general, using Lee and Strazicich (2003), we tend to reject more
H0 (unit root).

26
RS – EC2 - Lecture 16

Autoregressive Unit Root - Relevance

• We can always decompose a unit root process into the sum of a
random walk and a stable process. This is known as the Beveridge-
Nelson (1981) (BN) composition.

• Let yt ~ I(1), rt ~ RW and ct ~I(0).

yt = rt + ct.

Since ct is stable it has a Wold decomposition:

(1  L ) y t   ( L )  t
Then,
(1  L) yt  ( L) t  (1) t  (( L)  (1))t
 (1) t  ( L) *  t
where ψ(1)=0. Then,
yt  (1)(1  L) 1  t  ( L) * (1  L) 1  t  rt  ct

Autoregressive Unit Root - Relevance

• Usual finding in economics: Many time series seem to have unit
roots. But, there is debate over power of unit root tests and the effect
of structural breaks.

Example: Consumption, output, stock prices, interest rates,

unemployment, size, compensation are usually I(1).

• Sometimes a linear combination of I(1) series produces an I(0). For

example, (log consumption– log output) is stationary. This situation is
called cointegration.

• Practical problems with cointegration:

- Asymptotics change completely.
- Not enough data to definitively say we have cointegration.

27
RS – EC2 - Lecture 16

Autoregressive Unit Root – Structural Breaks 2

• Nelson and Plosser (1982) tested using ADF 14 macroeconomic
series (GNP, IP, employment, etc) for unit roots: Rejected H0 for only
one. Summary of results from tests allowing for structural breaks,
from Glyn et al. (2007):

28