See

discussions, stats, and author profiles for this publication at: http://www.researchgate.net/publication/235922660

Deterministic or Stochastic Trend: Decision on

the Basis of the Augmented Dickey-Fuller Test
ARTICLE in METHODOLOGY EUROPEAN JOURNAL OF RESEARCH METHODS FOR THE BEHAVIORAL AND SOCIAL
SCIENCES JANUARY 2010
Impact Factor: 0.33 DOI: 10.1027/1614-2241/a000009

CITATIONS

DOWNLOADS

VIEWS

2,587

530

1 AUTHOR:
Tatjana Stadnitski
Universitt Ulm
11 PUBLICATIONS 35 CITATIONS
SEE PROFILE

Available from: Tatjana Stadnitski

Retrieved on: 14 September 2015

Original Article

Deterministic or Stochastic Trend

Decision on the Basis of the Augmented
Dickey-Fuller Test
Tetiana Stadnytska
Department of Psychology, University of Heidelberg, Germany
Abstract. Time series with deterministic and stochastic trends possess different memory characteristics and exhibit dissimilar long-range
development. Trending series are nonstationary and must be transformed to be stabilized. The choice of correct transformation depends on
patterns of nonstationarity in the data. Inappropriate transformations are consequential for subsequent analysis and should be omitted. The
objectives of this article are (1) to introduce unit root testing procedures, (2) to evaluate the strategies for distinguishing between stochastic and
deterministic alternatives by means of Monte Carlo experiments, and (3) to demonstrate their implementation on empirical examples using SAS
for Windows.
Keywords: unit root tests, ADF test, time series, trend stationary, difference stationary, nonstationarity, integrated, Monte Carlo experiments

Nonstationary time series with a changing mean or variance

are common in psychology (Fortes, Ninot, & Delignie`res,
2004; Glass, Willson, & Gottman, 1975; Gottman, 1981;
McCleary & Hay, 1980; Ninot, Fortes, & Delignie`res,
2005; Velicer & Colby, 1997; Velicer & Fava, 2003;
Warner, 1998). Most psychological time series are unstable
due to a deterministic trend or stochastic drift. Nonstationary processes with deterministic and stochastic trend components possess different memory characteristics and
exhibit dissimilar long-range development. In social and
behavioral sciences, the goal of time-series analysis is usually to determine the nature of the process that describes an
observed behavior, to measure the effects of an intervention,
as in an interrupted time-series experiment, or to forecast
future values of the series under consideration. In the latter
two cases, stationarity of the series under study is required,
which implies some form of trend removal for nonstationary data. The choice of an appropriate detrending procedure
depends on the cause of nonstationarity. Mis-specifying the
trend characteristics of the data can result in biased tests and
false predictions (Ashley & Verbrugge, 2004, 2006;
Diebold & Kilian, 2000; Diebold & Senhadji, 1996;
Psaradakis & Sola, 2003). If two or more time series are
nonstationary due to a stochastic drift, the long-run equilibrium of such time series can be modeled by means of cointegration techniques. Hence, clarifying the cause of
nonstationarity in the data represents the rst step in the
cointegration modeling (Stroe-Kunold & Werner, 2007,
2008). This paper introduces unit root testing strategies
allowing one to distinguish between stochastic and deterministic alternatives, evaluate them by means of Monte
Carlo experiments, and demonstrate their implementation
on empirical examples.
2010 Hogrefe Publishing

Different Types of Nonstationarity

This part of the article introduces the concept of stationarity,
discusses the difference between stochastic and deterministic
trends, and presents unit root testing as a method for
distinguishing between stationary and nonstationary series.

Stationarity
A process is said to be stationary if its mean, variance, and
covariance do not change over time. If that is not the case,
we deal with a nonstationary process. If the goal of analysis
consists in measuring the effects of an intervention, as in an
interrupted time-series experiment, or in forecasting future
values of the series, stationarity (i.e., stability) of the data
under consideration is required.
A great deal of psychological time series has a timevarying mean, a time-varying variance, or both. For further
analysis, such time series must be transformed to make
them stationary. The transformation method depends on
the cause of nonstationarity. The consequences of a false
treatment can be rather serious; unfortunately, this is not
emphasized in the time-series textbooks used among psychologists. Some descriptions even suggest that two popular
methods for stabilizing nonstationary series, differencing
and ordinary least square regression, are interchangeable
and that the choice of transformation technique is simply
a matter of researchers preference (see e.g., Warner,
1998, p. 39). One of the objectives of this paper is to
emphasize the importance of proper stationarity transformation for empirical time-series research.
Methodology 2010; Vol. 6(2):8392
DOI: 10.1027/1614-2241/a000009

84

T. Stadnytska: Deterministic or Stochastic Trend

Yt = Yt1 + ut

Figure 1. Nonstationary processes and

their autocorrelation and partial autocorrelation functions: (a) pure random walk,
(b) random walk with drift, and (c) deterministic time trend.

ut IID(0, 2)

ACF

-2

1.0

1.0

0.5

0.5

0.0

-6

-1.0

-1.0
1 2 3 4 5 6 7 8 9 10

1 12 23 34 45 56 67 78 89 100

Yt = 2.0 + Yt1 + ut

0.0
-0.5

-0.5

-4

PACF

1 2 3 4 5 6 7 8 9 10

Lag

Lag

ut IID(0, 2)

30

15

ACF

20

10

1.0

1.0

0.5

0.5

0.0

-1.0

-1.0
1 12 23 34 45 56 67 78 89 100

1 2 3 4 5 6 7 8 9 10

Yt = 0.1t + 0.2Yt 1 + at

0.0
-0.5

-0.5

PACF

25

1 2 3 4 5 6 7 8 9 10

Lag

at = 0.5at 1 + u t

Lag

ut IID(0, 2)

12
1.0

0.5

0.5

ACF

3
0

PACF

1.0

0.0

-3

-1.0
1 12 23 34 45 56 67 78 89 100

0.0
-0.5

-0.5

-1.0
1 2 3 4 5 6 7 8 9 10

1 2 3 4 5 6 7 8 9 10

Lag

Lag

Nonstationary Processes

variance increase over time. Random walk processes are

nonstationary, but their rst differences

Figure 1 shows the three most common nonstationary processes and their autocorrelation and partial autocorrelation
functions (ACF and PACF, respectively). The process


1
Y t Y t1 ut with ut  IIDN 0; r2
is called a pure random walk. The mean of this process is
equal to its initial value but its variance (tr2) increases
indenitely over time. A pure random walk process can
also be represented as the sum of random shocks
X
ut :
2
Yt
As a result, the impact of a particular shock does not dissipate, and the random walk remembers the shock forever.
That is why a random walk is said to have an innite
memory. If a constant term is present in the equation
Y t a Y t1 ut ;

then Yt is called random walk with drift, where a is known

as the drift parameter. Depending on a being negative or
positive, Yt exhibits a negative or positive stochastic trend.
For a random walk with drift, the mean as well as the

Y t Y t  Y t1

are stationary. Hence, both types of random walks are

called difference stationary (DS) processes.
Random walk models are also known in the time-series
literature as unit root processes. A situation of nonstationarity is called the unit root problem, if in the rst-order autoregressive model
Y t qY t1 ut ;

where q is 1. The name unit root is due to the fact that

q = 1. (The autoregressive model can also be written as
1  LY t ut . The term unit root refers to the root of
the polynomial in the lag operator.1) Random walk is a
specic case of a more general class of stochastic models
known as integrated processes. An integrated process of
rst order is represented by an equation
Y t a Y t1 at ;

where any stationary autoregressive moving-average,

ARMA (p, q) process can generate the random part at.

Lag operator L: LYt = Yt1, L2Yt = Yt2 and so on. If (1  L) = 0, we obtain, L = 1, hence the name unit root.

Methodology 2010; Vol. 6(2):8392

2010 Hogrefe Publishing

T. Stadnytska: Deterministic or Stochastic Trend

Therefore, random walk processes are integrated of

order 1, denoted as I(1) or (0, 1, 0) in terms of Box and
Jenkins ARIMA methodology. In general, if a time series
has to be differenced d times to make it stationary, that
series is called integrated of the order d.
The development of the third process in Figure 1, represented by the equation
Y t 0:1t 0:2Y t1 at

with at 0:5at1 ut ;

is determined by a deterministic time trend. This process has

stable variance and a changing mean. In a more simple case
Y t b1 b2 t ut ;

the mean is, for instance, b1 + b2t. If we subtract this

mean from Yt, the resulting series will be stationary (this
procedure is known as polynomial detrending). That is
why processes of this type are called trend stationary
(TS). In contrast to integrated series, processes with a
deterministic trend do not exhibit an innite memory.
The deviations from the trend line do not contribute to
the long-run development of the series. In the case of a
stochastic trend, however, the random component at
affects the long-run course of the series.
Therefore, the proper transformation method of nonstationary data crucially depends on the data generating process (DGP). If an empirical time series is a realization of the
random walk process, the solution here is to take the rst differences of the time series. If a series is stationary around the
trend line, the correct way to transform such a time series is to
regress it on time. The residuals from this regression will then
be stationary. As Figure 1 shows, realizations of DS and TS
processes can appear very similar and it is especially difcult to decide whether the trend in a time series is stochastic
or deterministic (Cases b and c in Figure 1). As a result,
inappropriate transformations are not unusual in practice.
Different studies have shown that consequences of misspecifying the trend characteristics of the data can be rather
serious.

Consequences of Inappropriate
Transformations
Chan, Hayya, and Ord (1977) analyzed the effects of incorrect transformation on the autocorrelation and the power
spectral density functions. These authors showed that the
ACF of residuals from linear regression of a random walk
series on time are not stationary and tend to exhibit cycles
of increasing length and amplitude around a tted trend
line as sample size gets larger. That is why erroneous detrending of DS series is also called underdifferencing. The
residuals of inappropriately differenced TS series follow a
noninvertible moving-average process. This is known as
overdifferencing. There has been some debate in the literature arguing that overdifferencing is a less serious error
than underdifferencing (see Maddala & Kim, 1998, for an
overview).
Nelson and Kang (1981, 1984) detected articial periodicity in inappropriately detrended time series and presented
2010 Hogrefe Publishing

85

several ways in which investigators could obtain misleading

results using underdifferenced series. Diebold and Senhadji
(1996) showed that applying DS and TS models to the same
time series could result in very different predictions. SchenkHoppe` (2001) and Psaradakis and Sola (2003) demonstrated
dramatic consequences of inappropriate detrending within
the scope of business cycle research. In a recent paper,
Dagum and Giannerini (2006) investigated the effect of
erroneous transformations on tests detecting nonlinearity
and a unit root. The authors concluded that either wrong
differencing or wrong detrending leads to biased results.
Ashley and Verbrugge (2004, 2006) studied the effects of
false transformations in the context of ordinary parameter
inference in simple linear models. Underdifferencing yielded
seriously oversized tests. Overdifferencing in turn produced
very distorted estimated impulse response functions. Distortions became increasingly severe as sample size grew.
To summarize, memory characteristics and long-range
development of a time series crucially depend on whether
its trend component is deterministic or stochastic. A deterministic trend is completely predictable and not variable,
whereas a stochastic trend is not predictable. Series with a
stochastic trend have to be differenced to make them stationary. For series with a deterministic trend, polynomial detrending is a correct transformation to achieve stationarity.
Inappropriate transformation is consequential for subsequent
analysis and should be omitted. It is impossible to distinguish between stochastic and deterministic alternatives visually or through analyzing ACF and PACF of the series under
study.

Unit Root Testing

Testing for unit roots is a common practice in econometrics.
To most psychologists this methodology might be unfamiliar; therefore this part of the article briey presents the basic
concepts of the unit root approach and demonstrates the test
procedure on simulated data.
Augmented Dickey-Fuller Test
There exist numerous unit root tests. One of the most popular among them is the Augmented Dickey-Fuller (ADF)
test. (See Gujarati, 2003, for an introduction; see Maddala
& Kim, 1998, for an overview; consult Hamilton, 1994, at
the advanced level.) In the simplest case of an uncorrelated
error term the test begins by estimating the equation
Y t qY t1 ut :

For reasons of convenience, this equation is written in the

form

Y t q  1Y t1 ut dY t1 ut

Y t Y t  Y t1 :

with
10

The null hypothesis is d = 0, which means q = 1 or there

is a unit root, implying that the time series under consideration is integrated. Dickey and Fuller (1981) have shown
Methodology 2010; Vol. 6(2):8392

86

T. Stadnytska: Deterministic or Stochastic Trend

that under the null hypothesis, the estimated t value of d

follows the s statistics and computed the critical s values
on the basis of Monte Carlo simulations. The authors also
introduced a competing F test with the usual F computation but using special critical values. Elder and Kennedy
(2001a) showed in their recent paper, however, that in testing for stationarity, the t test is preferable.
The actual procedure of implementing the ADF test
involves several decisions. To allow for various possibilities
of nonstationarity, the ADF test is estimated in three different forms:

a large p is necessary (for further details see Lopez,

1997; Ng & Perron, 1995, 2001; Stock, 1994).
As the majority of the unit root procedures, the ADF
test suffers from size distortion through misspecication
of H0 or an inappropriate selection of lagged terms (the
true signicance level exceeds the nominal level). Another
drawback is the low power in smaller samples and in cases
of q near 1.
Example
The following SAS code simulates the series Y with
T = 100 yt = yt1 + at, where at = ut  0.6ut1 and
ut  IIDN (0, 1).

(I) Yt is a random walk DYt = dYt1 + ut;

(II) Yt is a random walk with drift DYt = a + dYt1 + ut;
(III) Yt is a random walk with drift around a deterministic
trend DYt = a + bt + dYt1 + ut.

data series;
y=0; y1=0; a=0; u=0; a1=0; teta=0.6;
keep t y;

In each case, the null hypothesis is d = 0. The alternative

hypothesis is that d < 0 (one-sided test); that is, the time
series is stationary. If the null hypothesis is rejected, it means
in the rst case that Yt is a stationary series with a zero mean.
In the second case, Yt is stationary with a nonzero mean and
in the third case, Yt is stationary around a deterministic trend.
The critical s values are different for each of the three preceding specications of the ADF test. If the error term in
the model is autocorrelated, the ADF test is conducted by
augmenting the preceding three equations by adding the
lagged values of the dependent variable DYt:
p
X

Y t a bt dY t1
ci
Y ti ut :
11

do t=50 to 100;
u=rannor(59837);
y=y1+a; a=uteta*a1;a1=u; y1=y;
if t gt 0 then output;
end;
run;
When testing for unit roots, it is crucial to specify the null
and alternative hypotheses appropriately. For example, if
the data are not growing, the hypotheses should reect this.
Therefore the rst step of analysis is to examine if the
observed series exhibits an increasing or decreasing trend.
Figure 2a shows that there is no apparent positive or negative trend in the generated series. Slow decay of the ACF
suggests that the process may be nonstationary; so we have
to decide between the H0 yt is I(1) without drift and the
H1 yt is I(0) with nonzero mean. Hence the test regression
is DYt = a + dYt1 + at.
An important practical issue for the correct implementation of the ADF test is the specication of the order of serial

i1

The number of lagged differenced terms (p) is determined

empirically using information criteria such as the Akaikes
information criterion (AIC) and the Bayesian information
criterion (BIC), the idea being to include enough terms so
that the error part at is serially uncorrelated. In general, a
small p is adequate for autoregressive errors or ARMA
processes with small moving-average components. For
error terms with moving-average coefcients near 1,

a 6.00

2.00

1.0

0.5

0.5

PACF

ACF

4.00

1.0

0.0

-1.0

-1.0
1

12 23 34 45 56 67 78 89 100

0.0
-0.5

-0.5

0.00

1 2 3 4 5 6 7 8 9 10 1112

Figure 2. Generated yt = yt1 +

at  0.6at1 + ut series and its
ACF and PACF: (a) nontransformed and (b) differenced
(d = 1).

1 2 3 4 5 6 7 8 9 10 1112

Lag

Lag

-2.00

1.0

1.0

0.5

0.5

PACF

0.00

ACF

Y(1)

2.00

0.0

-0.5

-0.5

-1.0

-1.0

-4.00
2 13 24 35 46 57 68 79 90

Methodology 2010; Vol. 6(2):8392

0.0

1 2 3 4 5 6 7 8 9 10 1112

Lag

1 2 3 4 5 6 7 8 9 10 1112

Lag

2010 Hogrefe Publishing

T. Stadnytska: Deterministic or Stochastic Trend

correlation (p) in the error term at. Ng and Perron (1995)

suggested the following lag length selection procedure.
First, set an upper bound pmax for p. Schwert (1989) proposed to use
" 
1=4 #
T
pmax INT 12
:
12
100
For series with length up to 100 observations, pmax = 8 is
usually sufcient. Next, estimate the ADF test regression
with p = pmax. If the absolute value of the t statistic for testing the signicance of the last lagged difference is > 1.6,
then set p = pmax and perform the unit root test. Otherwise
reduce the lag length by one and repeat the process.
The following SAS statements test the regression

87

Table 2. SAS output for the ADF unit root test for the
simulated yt = yt1 + at  0.6at1 + ut series
Type
Single mean

Lag

Pr < s

1.89

.3368

proc arima data= series;

identify var=y stationarity=(adf=(6));
run;
Table 2 presents the relevant parts of the SAS output.
With p = 6 the ADF t statistic is 1.89, the p value .3368.
Hence we cannot reject the unit root null hypothesis. Therefore the results of the ADF test correctly indicate the unit
root model without drift for the simulated series.

Y t a dY t1 c1
Y t1 c2
Y t2   

c8
Y t8 ut :

13

Deterministic Versus Stochastic Trend

data series;
set series;
y_dif=dif(y);
ly=lag(y);
dy1=lag(y_dif);
dy2=lag2(y_dif);
dy3=lag3(y_dif);
dy4=lag4(y_dif);
dy5=lag5(y_dif);
dy6=lag6(y_dif);
dy7=lag7(y_dif);
dy8=lag8(y_dif);
run;
proc reg data=series outest=est;
model y_dif= ly dy1 dy2 dy3 dy4 dy5 dy6
dy7 dy8/ aic bic;
run;
proc print data=est;
run;
Table 1 shows that the absolute value of the t statistic for
p = 8 is < 1.6, so the procedure is repeated for 7, 6, and
5 lags. Following the Ng-Perron backward selection algorithm and minimizing the AIC and the BIC, p = 6 lags were
chosen.
The following SAS code conducts the ADF test with
p = 6.
Table 1. Summary statistics for the Ng-Perron backward
selection procedure for the simulated yt = yt1 +
at  0.6at1 + ut series
Last lag
8
7
6
5

|t|

Pr(<|t|)

AIC

BIC

0.56
0.85
1.8
1.4

.5781
.3954
.0747
.3025

9.75
8.28
5.99
7.17

14.19
12.20
9.48
10.28

2010 Hogrefe Publishing

This part of the paper focuses on a special issue of growing

time series. As noted previously, such growth could occur
due to either a deterministic or stochastic trend. In the former case, there is no unit root, meaning the series under consideration is I(0) around a deterministic trend. In the latter
case, we have a unit root with drift or, in other words, the
I(1) process. The simultaneous existence of a unit root and
a deterministic trend is thought to be unrealistic (for explanations see Elder & Kennedy, 2001b; Holden & Perman,
1994; Perron, 1988). Therefore, the rst hypothesis from
the ADF family (a pure random walk) can be ruled out because it implies that the DGP is not growing. Hence we
must decide between two competing hypotheses: (1) there
is a unit root with drift d = 0 (q = 1), a 5 0 and (2) there
is a time trend without a unit root d < 0 (q < 1), b 5 0.
An informal testing strategy used among practitioners is
to reject a unit root when at least one of the ADF tests
rejects. In other words, the null hypothesis of the unit root
is dismissed if a test with an intercept does not reject the null
and one including an additional time trend term does (Kim,
Leybourne, & Newbold, 2004). Perron (1988) introduced a
sequential testing strategy and recommended starting from
the most general trend specication and testing down to
more restricted specications. Ayat and Burridge (2000) also
advocate a sequential testing starting with the highest trend
degree maintained.
Elder and Kennedy (2001b) propose beginning with the
hypothesis of a unit root with drift d = 0 (q = 1), a 5 0. If
this null is not rejected, the trend is stochastic. In case of
rejection, the series under study is probably stationary
around a deterministic time trend because two other possibilities (d 5 0, b = 0, and d = 0, b 5 0) are unreasonable
and can be ruled out. There is strong evidence, however, that
the last strategy works poorly. West (1987) demonstrated
that if a linear trend term is incorrectly omitted from the estimating equation, the rejection probability of H0 converges to
0. In other words, one almost never rejects a unit root
hypothesis. It is noteworthy that the inclusion of a redundant
Methodology 2010; Vol. 6(2):8392

88

T. Stadnytska: Deterministic or Stochastic Trend

time trend parameter leads to a considerable loss in power if

the series under consideration is stationary about an intercept
alone (Dickey, 1984). For our purposes, however, this case
is irrelevant because the H1 implies that the DGP is not
growing.
In sum, there are two possibilities for modeling long-run
growth: (1) a unit root implying a stochastic trend and (2) a
deterministic time trend. To differentiate between these two
alternatives, the Dickey-Fuller regression must include a
deterministic time trend term.

Monte Carlo Study

The following Monte Carlo experiments are designed to
demonstrate that employing Model III of the ADF family
as a major test regression allows distinguishing between
series with stochastic and deterministic trends.
Method
The evaluated strategy suggests starting with the most
general H0 of unit root with drift around a deterministic
trend (Model III) and then continuing testing with a more
restricted H0 of a unit root with drift component only
(Model II). If this strategy works, the ADF test for simulated
TS series will reject the third null hypothesis because the
series are stationary around a deterministic trend. In other
words, a small Type II error probability b or a high power
(1  b) of the ADF test is expected. Testing of the second
H0 serves as a check. No power at all is expected in this case
because the ADF test is known to treat the growth in the
series as a signature of the random walk (West, 1987). For
simulated random walks with drift, the correct test decision
is to maintain both null hypotheses because the series contain a unit root. Thus the frequency of the false rejection
of the null hypotheses (II) and (III) (i.e., test size) is expected
to correspond with the nominal Type I error level. In summation, for TS series different test powers are expected
and for DS series similar test sizes in two subtests.
The strategy was examined on 6 DS and 12 TS models.
The following independent variables were manipulated: (1)
type of model, (2) error term, and (3) number of lags. Series
with a stochastic trend were generated applying the following DGP: Yt = a + Yt1 + at with two different values of
the drift parameter a = .2 and .5. The DGP of series with
a deterministic trend was Yt = 0.1t + qYt1 + at with q
equal to 0, 0.2, 0.5, and 0.8 to account for different possible
degrees of serial dependency. To limit the scope of the study,
only three types of the error term at (white noise or
ARMA(0, 0), a rst-order autoregressive model with
/ = 0.5 or ARMA(1, 0) and a rst-order moving-average
model with h = 0.5 or ARMA(0, 1)) were used in DS
and TS cases. Such simple error structures are most common in practice. The results were computed for the rst
six lags to demonstrate the inuence of the lag length selection on the performance of the ADF test.
Each simulated time series consisted of 100 observations
and was replicated 1,000 times. Within the scope of the
Methodology 2010; Vol. 6(2):8392

ARIMA methodology, samples of 100 observations or more

represent an optimum length (Box & Pierce, 1970; Glass
et al., 1975; Granger & Newbold, 1986; Ljung & Box,
1978; McCleary & Hay, 1980; Velicer & Fava, 2003). For
effects of sample size on the performance of the ADF test,
refer to DeJong, Nankervis, Savin, and Whiteman (1992).
The percentage of signicant decisions of the ADF test
served as a dependent measurement.
All computations and the generation of independent
N(0, 1) innovations ut were performed with the SAS System
for Windows Version 9.1. For the ADF test, the following
statement of PROC ARIMA was used: identify var=name
stationarity= (adf=6). The 5% critical s values for the null
hypotheses d = 0, a 5 0 and d = 0, a 5 0, b 5 0 were
2.89 and 3.45, respectively.
Results and Discussion
Results obtained for models with a stochastic trend are considered rst. In this case, the DGP contains a unit root. Thus,
a correct test decision is to maintain the null hypotheses in
the both subtests. Table 3 presents the rejection percentages
of the ADF test at a nominal signicance level of 5%.
Although the inclusion of a redundant time trend parameter
(b) in the estimating equation enhances the number of false
test decisions, for all simulated models however the frequency of H0 rejections or the test size is close to the nominal level of signicance, provided a correct lag length is
selected. Recall that for models containing error terms with
a moving-average part, employing larger lags is appropriate.
This explains the values > 5% for lags 15 in the movingaverage case (0, 1). The obtained results are in accordance
with the evaluated testing strategy since we expect for DS
series the adherence to the nominal level of signicance
and no noticeable discrepancies in size of the two subtests.
Table 4 presents the percentages of signicant ADF tests
for models with a deterministic trend. In this case, the DGP
does not contain a unit root and therefore the rejection of H0
represents a correct test decision. It can be seen from the
left-hand section of Table 4 that omitting a time trend term
from the estimating regression leads to the lack of any test
power. In other words, there is no possibility to reject the
null hypothesis of a unit root. Including a deterministic
parameter, on the other hand, ensures quite good results.
This enormous discrepancy in the powers of the two subtests serves as distinct evidence in favor of the suggested
strategy. The right-hand section of Table 4 additionally
shows that the quality of test decisions is better for low q
values than for high q ones. As expected, the worst results
are obtained for q near 1. Recall that q = 1 implies a unit
root. For all TS models, the power of the ADF test decreases
with the number of lagged terms used.
The simulation results demonstrated that for growing
time series the test should begin with the third most general
hypothesis from the ADF family (a unit root with drift and a
time trend) and then continue with the more restricted case
of a unit root with drift. Rejection of the null hypothesis in
the rst test can be treated as strong evidence of a deterministic trend. If the null is not rejected in either test, the growth
2010 Hogrefe Publishing

T. Stadnytska: Deterministic or Stochastic Trend

89

Table 3. Percentage of signicant decisions of the ADF test at the nominal 5% level of signicance for DS series with
DGP: Yt = a + Yt1 + at; T = 100; 1,000 replications

at/Lag

Model II

Model III

H0: d = 0 (q = 1) a 5 0

H0: d = 0 (q = 1) a 5 0 b 5 0

a = .2
(0, 0)
(1, 0)
(0, 1)

2.6
4.2
0.2

2.4
4.0
0.8

2.0
4.2
0.7

2.3
4.7
0.6

3.0
5.1
0.8

2.7
4.7
0.9

4.5
5.5
29.0

5.7
5.6
12.0

5.0
5.5
7.8

4.6
4.5
6.2

4.5
5.1
6.1

5.3
5.5
4.7

a = .5
(0, 0)
(1, 0)
(0, 1)

0.6
1.9
0

0.2
1.6
0.4

0.4
1.7
0.5

0.6
2.1
0.3

0.8
2.5
0.5

0.8
2.3
0.2

4.5
5.5
29.0

5.7
5.6
12.0

5.0
5.5
7.8

4.6
4.5
6.2

4.5
5.1
6.1

5.3
5.5
4.7

Table 4. Percentage of signicant decisions of the ADF test at the nominal 5% level of signicance for TS series with
DGP: Yt = 0.1t + qYt1 + at; T = 100; 1,000 replications
Model II

Model III

H0: d = 0 (q = 1) a 5 0

H0: d = 0 (q = 1) a 5 0 b 5 0

at/Lag

q = 0.0
(0, 0)
(1, 0)
(0, 1)

0
0
0

0
0
0

0
0
0

0
0
0

0
0
0

0
0
0

100
99.9
99.9

100
97.2
99.0

99.1
89.3
98.0

96.4
75.7
89.9

86.4
62.6
79.1

69.0
47.3
62.7

q = 0.2
(0, 0)
(1, 0)
(0, 1)

0
0
0.2

0
0
0

0
0
0

0
0
0

0
0
0

0
0
0

100
99.5
99.9

100
94.5
98.2

98.4
84.6
96.6

92.1
69.6
84.4

80.4
57.4
72.8

63.7
43.4
56.9

q = 0.5
(0, 0)
(1, 0)
(0, 1)

0
1.9
2.8

0
0.2
0

0
0.3
0

0
0.1
0

0
0
0

0
0
0

99.9
92.0
99.7

97.2
78.9
87.1

89.3
68.2
86.4

75.7
52.1
65.9

62.6
43.6
56.8

47.3
32.6
44.8

q = 0.8
(0, 0)
(1, 0)
(0, 1)

0
11.5
11.9

0
7.9
2.2

0
6.0
2.8

0
4.9
1.9

0
4.5
2.7

0
3.5
1.3

56.8
38.6
76.2

41.8
30.7
28.0

35.6
27.2
39.5

29.8
21.4
24.8

24.0
18.6
25.8

18.0
15.7
20.9

in the observed series is probably due to a stochastic trend.

Furthermore, the ndings underlined the importance of the
proper choice of lag length in unit root testing. The presented analysis was limited to rather simple models. Using
more complex error structures or DGPs with nonlinear
trends could lead to divergent results (see Kim et al., 2004).

Empirical Demonstration
To illustrate the ADF test procedure for trend cases, timeseries raw data of trafc fatalities for New York State from
January 1951 to April 1960 were employed. The data are
2010 Hogrefe Publishing

freely available in the textbook of Glass et al. (1975).

Figure 3 shows the original series, its rst difference, and
the residuals from linear regression. Employing the
Box-Jenkins procedure, Glass and colleagues tted the
(0, 1, 1) model to the series. (For an elaborated strategy of
model selection combining different techniques, see
Stadnytska, Braun, and Werner (2008a, 2008b).) Therefore
according to Glass et al. (1975), a positive trend in the series
is due to a stochastic drift.
For a growing series in the ADF procedure the hypotheses to be tested are
H0: yt is I(1) with drift;
H1: yt is I(0) with deterministic time trend.
Methodology 2010; Vol. 6(2):8392

90

T. Stadnytska: Deterministic or Stochastic Trend

a 5.00
0.5

ACF

2.00

0.5

PACF

1.0

3.00

4.00

0.0

0.00

0.0
-0.5

-0.5

1.00

-1.0

-1.0
1 2 3 4 5 6 7 8 9 10 1112

1 12 23 34 45 56 67 78 89 100

Figure 3. New York state trafc

fatalities data from Glass et al.
(1975) and its ACF and PACF:
(a) nontransformed, (b) differenced, and (c) detrended.

1.0

1 2 3 4 5 6 7 8 9 10 11 12

Lag

Lag

1.0

1.0

0.00

0.5

0.5

-1.00

PACF

1.00

ACF

Y(1)

b 2.00

0.0

-2.00

-0.5

-3.00

-1.0
2 13 24 35 46 57 68 79 90

0.0
-0.5
-1.0
1 2 3 4 5 6 7 8 9 1011 12

1 2 3 4 5 6 7 8 9 1011 12

Lag

Lag

c 2.00

-1.00

1.0

1.0

0.5

0.5

PACF

0.00

ACF

Res

1.00

0.0
-0.5

-2.00

-0.5

-1.0
1 12 23 34 45 56 67 78 89 100

-1.0
1 2 3 4 5 6 7 8 9 101112

1 2 3 4 5 6 7 8 9 1011 12

Lag

Thus the test regression is
Y t a bt dY t1

p
P
ci
Y ti ut to capture the deterministic trend under
i1

the alternative. The number of lags in the test regression

can be determined using the Ng-Perron backward selection method. The following SAS code ts the test regression with p = 6 to the data and computes the AIC and
BIC.
data newyork;
set newyork;
y_dif=dif(y);
ly=lag(y);
dy1=lag(y_dif);
dy2=lag2(y_dif);
dy3=lag3(y_dif);
dy4=lag4(y_dif);
dy5=lag5(y_dif);
dy6=lag6(y_dif);
run;
proc reg data= newyork outest=est;
model y_dif=t ly dy1 dy2 dy3 dy4 dy5 dy6/
aic bic;
run;
proc print data=est;
run;

Methodology 2010; Vol. 6(2):8392

0.0

Lag

Table 5. Summary statistics for the Ng-Perron backward

selection procedure for the New York trafc
fatalities series
Last lag
6
5
4
3
2
1

|t|

AIC

BIC

RMSE

1.52
0.53
1.2
0.51
1.39
0.9

111.00
112.78
116.68
119.11
123.02
124.49

107.10
109.30
113.58
116.31
120.49
122.15

0.526
0.527
0.522
0.522
0.517
0.519

Table 5 summarizes the results from the Ng-Perron algorithm for lags 61. The results indicate that there is no need
of lagged differences in the test regression. (The absolute value of the t statistics is < 1.6 for all lags tested. The model
with p = 1 has the smallest AIC and BIC. Note that it does
not matter that the AIC or BIC values were all negative in
this example. The values are simply supposed to be compared to each other, meaning that we should pick the model
with the smallest actual (not absolute) value of AIC or BIC.)
The following SAS statements perform the ADF test for
lags 02.
proc arima data= newyork;
identify var=y stationarity=(adf=2);
run;

2010 Hogrefe Publishing

T. Stadnytska: Deterministic or Stochastic Trend

Table 6. SAS output for the ADF unit root test for the
New York trafc fatalities series
Lags

Pr < s

Single mean

0
1
2

6.07
3.39
3.04

< .0001
.0137
.0346

Trend

0
1
2

10.16
6.32
6.09

< .0001
< .0001
< .0001

Type

Table 6 shows the relevant parts of the SAS output. For

lags 02, the results strongly indicate the rejection of the null
hypotheses of unit root with drift. For instance, with p = 1
the ADF t statistic is 10.16 (p value < .0001). Therefore
in contrast to Glass et al. (1975), the ADF test indicates that
the series under study is TS.
This simple example illustrates the impact of model
selection on theoretical interpretation of phenomena under
study. For instance, the I(1) model identied by Glass and
colleagues suggests that the growth in the series is due to
accumulation of random shocks. According to this model,
an accidental increase in trafc fatalities in the past will persistently affect the future level of the series (since integrated
processes remember the shock forever). In contrast, the TS
model identied by means of the ADF test strategy does
not assume an inuence of random shocks on the longrun development of the series and explains the observed
growth as a function of the deterministic time trend (i.e.,
more cars, more accidents, and more fatalities).
Further, comparison of forecasts for the last observation
(t = 100) in the data demonstrated the superiority of the TS
model over the I(1) alternative. The forecast Dy100 = 0.87
(0.56) from the TS model:

Y t 2:18 0:018t 1:04Y t1 ut

14

lay close to the observed time-series value Dy100 = 0.94;

whereas the predicted value Dy100 = 0.23 (0.52) from
the (0, 1, 1) model:

Y t 0:017 ut  0:98ut1

15

deviated strongly from the true one.

Summary and Conclusions

Many psychological time series are nonstationary. Increasing or decreasing behavior of the observed series can be
due to a deterministic or stochastic trend. A deterministic
trend is completely predictable and not variable; a stochastic
trend is not predictable. Therefore, DS and TS models of the
same time series imply different memory properties of the
process under investigation. They also require different trend
removal procedures and may result in divergent predictions. Hence the decision which model to use is tremendously important for applied researchers. Furthermore, the
2010 Hogrefe Publishing

91

described unit root testing strategies play an important role

within the multivariate cointegration framework.
The ADF test allows determining the cause of nonstationarity in the data. Testing for a unit root always indicates
a testing strategy and not mere calculation of a single test
statistic. The rst step in the strategy is to plot data against
time and to rule out unreasonable hypotheses on the basis of
theoretical considerations (see Elder & Kennedy, 2001b).
For growing time series, the test regression should include
a constant and time trend (Hypothesis III from the ADF
family). In series with autocorrelated errors, the number of
lagged difference terms (p) used to approximate the ARMA
structure of the errors exercises a strong inuence on test
decisions. If p is too small, then the remaining serial correlation in the errors will bias the test. Including too many
lagged terms reduces the power of the test. Employing the
lag length selection procedure suggested by Ng and Perron
(1995) results in stable size of the ADF test with minimal
power loss.

References
Ashley, R., & Verbrugge, R. J. (2004). To difference or not to
difference: A Monte Carlo investigation of inference in vector
autoregressive models. VPI Economic Department Working
Paper. Retrieved from http://ashleymac.econ.vt.edu/working_
papers/9915.pdf.
Ashley, R., & Verbrugge, R. J. (2006). Comments on A critical
investigation on detrending procedures for non-linear processes. Journal of Macroeconomics, 28, 192194.
Ayat, L., & Burridge, P. (2000). Unit root tests in the presence of
uncertainty about the non-stochastic trend. Journal of
Econometrics, 95, 7196.
Box, G. E. P., & Pierce, W. A. (1970). Distribution of residual
autocorrelations in autoregressive-integrated moving average
time series models. Journal of American Statistical Association, 65, 15091526.
Chan, K. H., Hayya, J. C., & Ord, J. K. (1977). A note on trend
removal methods: The case of polynomial regression versus
variate differencing. Econometrica, 45, 737744.
Dagum, E. B., & Giannerini, S. (2006). A critical investigation
on detrending procedures for non-linear processes. Journal of
Macroeconomics, 28, 175191.
DeJong, D. N., Nankervis, J. C., Savin, N. E., & Whiteman, H.
(1992). The power problems of unit root tests in time series
with autoregressive errors. Journal of Econometrics, 53,
323343.
Dickey, D. A. (1984). Power of unit root tests. Proceedings of
Business and Economic Statistics Section of ASA, pp. 489
493.
Dickey, D. A., & Fuller, W. A. (1981). Likelihood ratio statistics
for autoregressive time series with a unit root. Econometrica,
49, 10571072.
Diebold, F. X., & Kilian, L. (2000). Unit root tests are useful for
selecting forecasting model. Journal of Business and
Economic Statistics, 18, 265273.
Diebold, F. X., & Senhadji, A. S. (1996). The uncertain unit root
in real GNP: Comment. American Economic Review, 86,
12911298.
Elder, J., & Kennedy, P. E. (2001a). F versus t test for unit roots.
Economic Bulletin, 3, 16.
Elder, J., & Kennedy, P. E. (2001b). Testing for unit roots: What
should students be taught? Journal of Economic Education,
32, 137146.
Methodology 2010; Vol. 6(2):8392

92

T. Stadnytska: Deterministic or Stochastic Trend

Fortes, M., Ninot, G., & Delignie`res, D. (2004). The dynamics of

self-esteem and physical self: Between preservation and
adaptation. Quality and Quantity, 38, 735751.
Glass, G. V., Willson, V. L., & Gottman, J. M. (1975). Design
and analysis of time-series experiments. Boulder, CO:
Colorado Associated University Press.
Gottman, J. M. (1981). Time-series analysis. New York:
Cambridge University Press.
Granger, C. W. J., & Newbold, P. (1986). Forecasting economic
time series. San Diego, CA: Academic Press.
Gujarati, D. (2003). Basic econometrics (4th ed.). New York:
McGraw-Hill.
Hamilton, J. D. (1994). Time series analysis. Princeton: Princeton University Press.
Holden, D., & Perman, R. (1994). Unit roots and cointegration
for the economist. In B. B. Rao (Ed.), Cointegration for the
applied economist (pp. 47112). New York: St. Martins.
Kim, T., Leybourne, S. J., & Newbold, P. (2004). Behaviour of
Dickey-Fuller unit root tests under trend misspecication.
Journal of Time Series Analysis, 25, 755764.
Ljung, G. M., & Box, G. E. P. (1978). On a measure of lack of
t in time series models. Biometrika, 65, 297303.
Lopez, J. H. (1997). The power of the ADF test. Economics
Letters, 57, 510.
Maddala, G. S., & Kim, I. (1998). Unit roots, cointegration and
structural change. Cambridge: University Press.
McCleary, R., & Hay, R. A. Jr. (1980). Applied time series
analysis for the social sciences. Beverly Hills: Sage.
Nelson, C. R., & Kang, H. (1981). Spurious periodicity in
inappropriately detrended time teries. Econometrica, 49,
741751.
Nelson, C. R., & Kang, H. (1984). Pitfalls in the use of time as
an explanatory variable in regression. Journal of Business
and Economic Statistics, 2, 7382.
Ng, S., & Perron, P. (1995). Unit root tests in ARMA models
with data dependent methods for the selection of the
truncation lag. Journal of the American Statistical Association, 90, 268281.
Ng, S., & Perron, P. (2001). Lag length selection and the
construction of the unit root test with good size and power.
Econometrica, 69, 15191554.
Ninot, G., Fortes, M., & Delignie`res, D. (2005). The dynamics of
self-esteem in adults over a six-month period: An exploratory
study. The Journal of Psychology, 139, 315330.
Perron, P. (1988). Trends and random walks in macroeconomic
time series. Journal of Economic Dynamics and Control, 12,
297332.
Psaradakis, Z., & Sola, M. (2003). On detrending and cyclical
asymmetry. Journal of Applied Econometrics, 18, 271289.
Schenk-Hoppe`, K. (2001). Economic growth and business
cycles: A critical comment on detrending time series. Studies
in Nonlinear Dynamics and Econometrics, 5, 7586.

Methodology 2010; Vol. 6(2):8392

Schwert, G. W. (1989). Tests for unit roots: A Monte Carlo

investigation. Journal of Business and Economic Statistics, 7,
147160.
Stadnytska, T., Braun, S., & Werner, J. (2008a). Model
identication of integrated ARMA processes. Multivariate
Behavior Research, 43, 128.
Stadnytska, T., Braun, S., & Werner, J. (2008b). Comparison of
automated procedures for ARMA model identication.
Behavior Research Methods, 40, 250262.
Stock, J. H. (1994). Unit roots, structural breaks, and trends. In
R. Engle & D. McFadden (Eds.), Handbook of Econometrics
(Vol. IV, pp. 27402843). Amsterdam: Elsevier.
Stroe-Kunold, E., & Werner, J. (2007). Sind psychologische
Prozesse kointegriert? Standortbestimmung und Perspektiven
der Kointegrationsmethodologie in der psychologischen
Forschung [Are psychological processes cointegrated? Present role and future perspectives of cointegration methodology
in psychological research]. Psychologische Rundschau, 58(4),
225237.
Stroe-Kunold, E., & Werner, J. (2008). Modeling human
dynamics by means of cointegration methodology. Methodology, 4(3), 113131.
Velicer, W. F., & Colby, S. M. (1997). Time series analysis for
prevention and treatment research. In K. J. Bryant, M.
Windle, & S. G. West (Eds.), The science of prevention:
Methodological advances from alcohol and substance
abuse research (pp. 211249). Washington, DC: American
Psychological Association.
Velicer, W. F., & Fava, J. l. (2003). Time series analysis. In
J. Schinka & W. F. Velicer (Eds.), Research methods in
psychology (pp. 581606). New York: Wiley.
Warner, R. M. (1998). Spectral analysis of time-series data. New
York: Guilford.
West, K. D. (1987). A note of the power of least squares tests for
a unit root. Economics Letters, 24, 249252.

Tetiana Stadnytska
Department of Psychology
University of Heidelberg
Hauptstrasse 47-51
69117 Heidelberg
Germany
Tel. +49 6221 547345
E-mail Tetiana.Stadnytska@psychologie.uni-heidelberg.de

2010 Hogrefe Publishing