Unit root,

differencing the time series, unit root test

(ADF test)
Beáta Stehlíková
Time series analysis

Unit root,differencing the time series, unit root test (ADF test) – p.1/27
Outline
• What is a unit root and what is its consequence
• If we have unit root - how to transform the data, so that
we can use the ARMA methodology
• How to find from the data that there is a unit root →
unit root tests

Unit root,differencing the time series, unit root test (ADF test) – p.2/27
Examples
• Consider the process yt = yt−1 + ut :
⋄ it is a nonstationary AR(1) process with a unit root
⋄ for its differences ∆yt = yt − yt−1 we have
∆yt = ut
⋄ so ∆yt is a stationary process
• Consider a nonstationary process with a unit root
1 1
(1 − L)(1 − L)xt = 1 + (1 − L)ut
2 3
Then for the differences
∆yt = yt − yt−1 = (1 − L)yt
we have
1 1
(1 − L)∆yt = 1 + (1 − L)ut ,
2 3
so ∆yt is a stationary process
Unit root,differencing the time series, unit root test (ADF test) – p.3/27
Examples
• Consider nonstationary process with a double unit root
1 2 1
(1 − L)(1 − L) xt = 1 + (1 − L)ut
2 3
The for the second differences
∆2 yt = ∆(∆yt ) = (1 − L)(1 − L)yt = (1 − L)2 yt
we have
1 2 1
(1 − L)∆ yt = 1 + (1 − L)ut ,
2 3
so ∆2 yt is a stationary process
• In general:
If the multiplicity of the unit root is k (and the others
are outsider the unit circle), then its k-th differences are
stationary.

Unit root,differencing the time series, unit root test (ADF test) – p.4/27
ARIMA models
• We need need to differentiate the process k times, in
order to get a stationary process, it is called intergrated
process of order k , denoted by I(k)
• If these k -th differences follow ARMA(p,q), then we
say that the original time series is ARIMA(p,k,q).
• For example xt , if
1 2 1
(1 − L)(1 − L) xt = 1 + (1 − L)ut ,
2 3
is ARIMA(1,2,1) process.

Unit root,differencing the time series, unit root test (ADF test) – p.5/27
Our aim
• Consider firstly AR(1) model: xt = δ + ρxt−1 + ut
• We want to:
⋄ test a hypothesis about a unit root (then, the
process is nonstationary), so H0 : ρ = 1
⋄ find out it can be rejected in favour of the
stationarity - H1 : ρ < 1

Unit root,differencing the time series, unit root test (ADF test) – p.6/27
Unit root and the t-statistics
We try to use the hypothesis testing about coefficients of a
regression model known from econometrics.
For example:
• Set x=1:200
• Simulate y=x+rnorm(200)*sigma
• Estimate the model y = c + ρ x + ε
• We note:
⋄ estimate of the parameter ρ
⋄ value of the t-statistics corresponding to hypothesis
H0 : ρ = 1 (which holds)
• Repeat 105 times and plot the histogram

Unit root,differencing the time series, unit root test (ADF test) – p.7/27
Unit root and the t-statistics
• Example of the simulated data:

Unit root,differencing the time series, unit root test (ADF test) – p.8/27
Unit root and the t-statistics
• Estimated regression:

• Estimated coefficient ρ is 0.95689

0.95689−1
• T-statistics to test ρ = 1 is 0.03459

Unit root,differencing the time series, unit root test (ADF test) – p.9/27
Unit root and the t-statistics
• Estimated of ρ: normal distribution

Unit root,differencing the time series, unit root test (ADF test) – p.10/27
Unit root and the t-statistics
• t-statistics to test H0 : ρ = 1: Student t-distribution

Unit root,differencing the time series, unit root test (ADF test) – p.11/27
Unit root and the t-statistics
Another simulation:
• Consider vector z generated as zt = zt−1 + εt
• Take x=z[1:200], y=z[2:201], so yt = zt , xt = zt−1
• Estimate the model y = c + ρ x + ε
• Note again:
⋄ estimate of parameter ρ
⋄ value of t-statistics to test H0 : ρ = 1 (which holds)
• Repeat 105 times and plot the histogram

Unit root,differencing the time series, unit root test (ADF test) – p.12/27
Unit root and the t-statistics
• Example of simulated data - time series s z :

Unit root,differencing the time series, unit root test (ADF test) – p.13/27
Unit root and the t-statistics
• Example of simulated data - data for regression

Unit root,differencing the time series, unit root test (ADF test) – p.14/27
Unit root and the t-statistics
• Estimated regression:

• Estimated coefficient ρ is 0.91962

0.91962−1
• t-statistics to test ρ = 1 is 0.02790 = −2.88

Unit root,differencing the time series, unit root test (ADF test) – p.15/27
Unit root and the t-statistics
• Estimates of parameter ρ: not a normal distribution

Unit root,differencing the time series, unit root test (ADF test) – p.16/27
Unit root and the t-statistics
• "t-statistics" to test H0 : ρ = 1: does not have a
t-distribution

• We cannot use critical values of the t-distribution

Unit root,differencing the time series, unit root test (ADF test) – p.17/27
Unit root and the t-statistics
• Solution - the mail idea:
⋄ we keep the test statistics
⋄ but we use different critical values
• Quantiles from our simulations:

- critical values should be similar

• Question:
⋄ What if we have a process of a higher order instead
of AR(1)?

Unit root,differencing the time series, unit root test (ADF test) – p.18/27
Unit root tests
• AR(1) process:
(1) yt = ρyt−1 + ut
unit root means that ρ = 1.
• Equivalently:
∆yt = (ρ − 1)yt−1 + ut
and we are interested in t-statistics from the
significance of coefficient at yt−1 - but with another
critical value
• This critical value
⋄ depends on number of data
⋄ changes, if equation (1) has a constant term or a
linear drift
• In general: ∆yt = α + βt + (ρ − 1)yt−1 + ut
Unit root,differencing the time series, unit root test (ADF test) – p.19/27
Unit root tests
• AR(p) process: yt = α1 yt−1 + α2 yt−2 + . . . αp yt−p + ut
unit root → α1 + . . . αp = 1.
• Write in the form:

yt = ρyt−1 +θ1 ∆yt−1 +θ2 ∆yt−2 +. . . +θp−1 ∆yt−p+1 +ut ,

Pp Pp
where ρ = j=1 αj , θi =− j=i+1 αj (i = 1, . . . , p − 1)
• Equivalently:
∆yt = (ρ−1)yt−1 +θ1 ∆yt−1 +θ2 ∆yt−2 +. . .+θp−1 ∆yt−p+1 +ut
and we are interested in t-statistics for coefficient at
yt−1
• In general: y can have a trend and/or intercept ⇒
∆yt = α + βt + (ρ − 1)yt−1 + θ1 ∆yt−1 + θ2 ∆yt−2 + . . .
+θp−1 ∆yt−p+1 + ut
Unit root,differencing the time series, unit root test (ADF test) – p.20/27
Augmented Dickey-Fuller test (ADF)
Wayne A. Fuller (1976)
David A. Dickey, Wayne A. Fuller (1979, 1981)

• We estimate
∆yt = α + βt + (ρ − 1)yt−1 + θ1 ∆yt−1 + . . . + θk ∆yt−k + ut
where we have to
⋄ decide whether to include a constant α and/or
linear trend β (depending on whether they are
present in process y )
⋄ choose k
• Then, we are interested in t-statistics from significance
test about coefficient in front of yt−1 , but with correct
critical values

Unit root,differencing the time series, unit root test (ADF test) – p.21/27
ADF test - critical values
• James G. MacKinnon (1991) - available as a part of an
expanded version from 2010:
James G. MacKinnon: Critical Values for Cointegration Tests. Queen’s Economics
Department Working Paper No. 1227, 2010..
Dostupné online: http://ideas.repec.org/p/qed/wpaper/1227.html

• Values obtaines by simulations:

Unit root,differencing the time series, unit root test (ADF test) – p.22/27
ADF test - critical values
• If we use T data points in the regression, the critical
value is β∞ + β1 /T + β2 /T 2
• In our example from the simulations: constant without
trend, T = 200:

⋄ for 1 percent:
−3.4336 − 5.999/200 − 29.25/2002 = −3.451
⋄ for 5 percent:
−2.8621 − 2.738/200 − 8.36/2002 = −2.879
• Compare with t-distribution (different) and quantiles
from simulations (ok)
Unit root,differencing the time series, unit root test (ADF test) – p.23/27
ADF test in R
• Library urca
• Function ur.df (ur - unit root, df - Dickey-Fuller) with
parameters:
⋄ type: possible values are drift (constant without
linear trend), trend (constant and linear trend),
none (nothing)
⋄ lags: maximal number of lags
⋄ selectlags: criterion for the choice of lags
(information criteri: AIC, BIC)

Unit root,differencing the time series, unit root test (ADF test) – p.24/27
ADF test in R
• Example: spread from the earlier lectures difference
between long-term and short-term rates)
• In R:
summary(ur.df(spread,type="drift",lags=8,selectlags="BIC")
• summary in order to obtain also critical values, not
only the test statistics

Unit root,differencing the time series, unit root test (ADF test) – p.25/27
ADF test in R
• Výstup: estimated regression and the test statistics:

Unit root,differencing the time series, unit root test (ADF test) – p.26/27
ADF test in R
• Output: test statistics and critical values:

• Criterion: Hypothesis about the unit root is rejected, if

the statistics is less than the critical value

Unit root,differencing the time series, unit root test (ADF test) – p.27/27