TESTING TIME SERIES DATA FOR STATIONARITY

Rizwan Mushtaq

[email protected]

Electronic
Electroniccopy
copyavailable
availableat:
at:https://ssrn.com/abstract=1911068
http://ssrn.com/abstract=1911068
Why Testing for Stationarity is Important

Testing data for stationarity is very important in research where the underlying variables based

on time. Moreover time series data analysis has many applications in many areas including

studying the relationship between wages and house prices, profits and dividends, and

consumption and GDP. An important econometric task is determining the most appropriate form

of the trend in the data. Many economic and financial time series exhibit trending behavior or

non-stationarity in the mean. Leading examples are asset prices, exchange rates and the levels of

macroeconomic aggregates like real GDP. In the beginning of the decade 1970s there was a great

debate about this topic. Granger and Newbold (1974) were the researchers, who give the idea

that the macroeconomic data as a rule contained stochastic trends, and this data is characterized

by unit root, they also suggest that using these variables in econometric models may lead towards

spurious regressions. So testing for stationarity is very important because the whole results of the

regression might be fabricated. In simple words we can say that trended series is called non-

stationary and with unit root and on the other hand non-trended series is a stationary series

characterized by without unit root. In formal way the series is called stationary if it satisfies three

conditions, otherwise it will be a non-stationary series.

i. Mean of Yt (E(Yt) remain same over time or time invariant. i.e.

E(Yt) = u,  t

Where the symbol , is use for all and (u) is any scalar

ii Variance of Yt (V (Yt) is time invariant. i.e.

V(Yt) = 2 ,  t

Electronic
Electroniccopy
copyavailable
availableat:
at:https://ssrn.com/abstract=1911068
http://ssrn.com/abstract=1911068
 iii Cov of Yt and Yt-s (cov (Yt, Yt-s) is time invariant, but can be depend upon the lag

length. i.e Cov( Yt, Yt-s) = s

If the above conditions do not hold series is non-stationary. In stationary series there is no link

between previous values. While the regression on trended (non-stationary) variables, is

meaningless, and provides misleading and spurious results. Normally most of the time series data

gives us meaningless results until appropriate econometrics and statistical tools were not applied.

Now the question is that how to know the data is stationary or non-stationary. We use certain

formal and informal tests as well to test the data for unit root. By plotting data into scatter plot

graph stationary of the data can be checked. The graphical method called informal way to check

the stationary while using E-views or other packages we can check the data for stationarity in a

formal method.

There are different formal methods use to check the data for stationarity proposed by different

researchers. But in this article our emphasis will endure on the test proposed by Dickey and

Fuller (1979, 1981) they develop a formal test for stationarity. The significant thing in their test

was that, testing for non-stationarity is equivalent to testing for the existence of unit root.

Moreover critique is also there on Dickey and Fuller test that the power of the test is very low,

about 30% it makes correct decisions. It is not considered an appropriate test and the entire tests

used for unit root have low power.

In this article Household final consumption expenditure per capita1 and GDP per capita2 of India

are used to test for stationarity. The data about these indicators covering the period of 1980-2009

1
Household final consumption expenditure per capita (constant 2000 US$)
2
GDP per capita, PPP (constant 2005 international $)

Electronic copy available at: https://ssrn.com/abstract=1911068

was congregated from World Bank official web site3. Primarily two methods are used to test

the data for stationarity the first one is informal method, based on simple graphical

representation of the series and the latter is the formal method including Dickey and Fuller and

Augmented Dickey and Fuller test using E-views.

1. Informal Method to Test the Data for Stationarity

With the help of Microsoft Excel the series of household final consumption is plotted. It is self-

explanatory chart, it shows increasing trend in GDP per capita of India and this series might

professed as non-stationary series and it has a unit root. Graph: 1-1 is presented below:

Graph: 1-1 GDP per capita, PPP (constant 2005 international $)

GDP
3500

3000

2500

2000

1500 GDP

1000

500

0
0 5 10 15 20 25 30 35

It might be possible to say that India is a developing country, and also a huge market for

multinationals. It is the second largest country of the world with reference to population, with

high birth rate. So it is expected increasing trends in GDP and consumption patterns.

3
http://data.worldbank.org/indicator

Electronic copy available at: https://ssrn.com/abstract=1911068

Graph: 1-2 Household final consumption expenditure per capita (constant 2000 US$)

CONSUMPTION
500
450
400
350
300
250
CONS
200
150
100
50
0
0 5 10 15 20 25 30 35

As we have noted earlier the series of GDP per capita showed increasing trends here graph 1-2

the series of household final consumption per capita also represents the non-stationary series, and

exposes the increasing trend as discussed earlier.

2. Dickey and Fuller Test (Formal Test for Stationarity)

This section encompassed a formal test of stationarity i.e. Dickey and Fuller test. This test

examine the null hypothesis of an autoregressive integrated moving average (ARIMA) against

stationary and alternatively. Some basic information about this test is provided in first section;

here we will discuss the methodology they use, and then carry on towards our real data analysis.

The formal version of Dickey Fuller test is explained here:

Electronic copy available at: https://ssrn.com/abstract=1911068

Consider an AR (1) model: Yt =  Yt-1 + ……………………………………….. (1)

Dickey Fuller suggest an alternative equation by subtracting Yt-1 from both sides of equation (1)

Yt – Yt-1=  Yt-1 – Yt-1+ 

Δ Yt = (-1) Yt-1+ 

Δ Yt =  Yt-1+ …………………………………...… (2)

Equation two is without constant where,  = -1. Dickey and Fuller also suggest two alternate

forms:

Constant only: Δ Yt =  +  Yt-1+ ………………………………….. (3)

Constant and Time Trend: Δ Yt =  + t +  Yt-1 + …………………………….. (4)

Testing the hypothesis

Ho: =0

H1: ˂0

Dickey-Fuller test with intercept was applied on both series to test the data for stationarity. The

null hypothesis is tested via t-statistics which is given by this formula:

………………………………………………….. (5)

If t calculated is greater than the critical value we do not reject our null hypothesis. In this

situation the variable under consideration will be non-stationary and has a unit root. On the other

hand if t calculated is less than the critical value we reject our null hypothesis. In this case the

Electronic copy available at: https://ssrn.com/abstract=1911068

underlying series would be a stationary series and it does not have the unit root. Firs the series is

tested on level if it does not become stationary than we travel further and test the series at first

and second difference sequentially. There is another method to reject or not to reject the null

hypothesis if the calculated value is on the right side of the critical value, on one sided tail (see

figure2-1 in appendix) we do not reject null hypothesis and if the calculated value is on the left

side of the critical value we reject the null hypothesis and conclude that the series does not has

unit root. P-value is also used to reject or accept the null hypothesis if the p-value < .05 reject

null hypothesis and vice versa.

Table: 2-1, Dickey Fuller Test at Level depicts that the calculated t-value is greater than critical

values at 1%, 5% and 10% significant levels. At levels both the underlying series are non-

stationary. Hence we do not reject the null hypothesis and accept alternate hypothesis that the

series has a unit root. Table: 2-2, named as Dickey Fuller Test at first difference provide similar

type of results as in table: 2-1, DF test statistics-GDP and Consumption series are also non-

stationary at first difference. Both the series are trended at first difference too, increasing trend

was also found via scatter plot graph, as this method refer as the informal way to test the

stationarity of the data.

Table: 2-1 Dickey-Fuller Test at Level

Variables t-statistics Level of significance Critical value

1% Critical Value -2.653401

DF test statistic-GDP 0.184098
5% Critical Value -1.953858

10% Critical Value -1.609571

1% Critical Value -2.653401

DF test statistic-Cons 0.062812 5% Critical Value -1.953858

Electronic copy available at: https://ssrn.com/abstract=1911068

 10% Critical Value -1.609571

Table: 2-2 Dickey-Fuller Test at First Difference

Variables t-statistics Level of significance Critical value

1% Critical Value -2.653401

DF test statistic-GDP -0.673457
5% Critical Value -1.953858

10% Critical Value -1.609571

1% Critical Value -2.653401

DF test statistic-Cons -0.409076 5% Critical Value -1.953858

10% Critical Value -1.609571

Table: 2-3, Dickey Fuller Test at Second Difference reveals that both the series are stationary at

second difference. Calculated t-value of GDP is (-9.128370) it less than the critical values at all

significant levels. Similarly t-statistics of consumption series is (-8.512923) it is also less than

the critical values at all significant levels. Therefore we reject our null hypothesis and conclude

that both the series are stationary at 1% significant level at second difference, and both the

series does not have the unit root. Our informal method to test stationarity also confirms the

results of formal test i.e. Dickey Fuller test. Graphs of both the series at second difference do

not demonstrate any kind of trend; there are fluctuations in the graphs. These fluctuations

epitomize the stationary of underlying economic series.

Table: 2-3 Dickey-Fuller Test at Second Difference

Variables t-statistics Level of significance Critical value

1% Critical Value -2.653401

DF test statistic-GDP -9.128370
5% Critical Value -1.953858

Electronic copy available at: https://ssrn.com/abstract=1911068

 10% Critical Value -1.609571

1% Critical Value -2.653401

DF test statistic-Cons -8.512923 5% Critical Value -1.953858

10% Critical Value -1.609571

3. Augmented Dickey-Fuller Test

Augmented Dickey Fuller Test is the extended version of simple Dickey Fuller test. Because of

the error term unlikely to be white noise4. They extended their test by including extra lagged in

terms of the dependent variables in order to eliminate the problem of autocorrelation. Normally

we use Augmented Dickey-Fuller test instead of simple Dickey-Fuller test. In simple words by

including the lagged values of dependent variable to the existing model, and continue this

procedure up till where the autocorrelation eliminated. It can be illustrated as:

Yt = 1 + 2 Yt + t…………………………………………………………………………….. (6)

Yt = 1 + 2 Yt + 3 Yt-1 + t……………………………………..……………………………. (7)

Yt = 1 + 2 Yt + 3 Yt-1 + 4 Yt-2 + t……………………………..………………………….. (8)

Now,

ΔYt =  Yt-1 + 1 ΔYt-1 + t……………………………………………………………………. (9)

ΔYt =  Yt-1 + 1 ΔYt-1 + 2 ΔYt-2…………………………………… + p ΔYt-p + t……………………. (10)

Continue this process up till, where autocorrelation eliminated. This expression could be written
as:

4
It is actually related with stationarity, except one thing. ”The third condition of stationary series, cov(Yt , Yt-1 = 0 )”
IID, Identical Independence Distribution. There is a slight difference between stationary and IID.

Electronic copy available at: https://ssrn.com/abstract=1911068

  ∑ 

  ∑ 

   ∑ 

Some common assumptions of ordinary least square (OLS) are discussed here:

1.  must be independent

2. There should be no heteroskedasticity5, should homogeneity.

3. There should no structural break, co-efficient should stable.

4. Error term should be normally distributed.

Testing for stationarity in Augmented Dickey-Fuller test follows the same procedure as in simple

Dickey-Fuller test. First, stationary is checked at level than at first difference and finally on

second difference. At first difference the equation will be as follows:

 ∑ 

  ∑ 

   ∑ 

5
Normally heteroskedasticity is defined as unequal spread. In econometrics the measure we usually use for spread is
the variance, and therefore heteroskedasticity deals with unequal variance.

10

Electronic copy available at: https://ssrn.com/abstract=1911068

If series is still not-stationary we can use same equations by replacing6 Δ2 with Δ3 rest process

will be same.

Testing for hypothesis using ADF

H0 :  = 0
H1 :  < 0

Table: 3-1 demonstrates the output of Augmented Dickey Fuller test at Levels. Variables of

concern are tested for stationary via E-Views. At levels both the variables are non-stationary as

the t-statistics of GDP is (7.645804) and the t-value of consumption is (8.198867), these values

are much higher than the critical values at all significant levels. So we do not reject our null

hypothesis and conclude that both the series has a unit root. Simple Dickey-Fuller test also

provide same kind of results at levels.

Table: 3-1 Augmented Dickey Fuller Test at Level

Variables t-statistics Level of significance Critical value

1% Critical Value -3.679322

ADF test statistic-GDP 7.645804
5% Critical Value -2.967767

10% Critical Value -2.622989

1% Critical Value -3.679322

ADF test statistic-Cons 8.198867 5% Critical Value -2.967767

10% Critical Value -2.622989

Now the series are not stationary at level, we continue our process to ADF at first difference.

Table: 3-2 denotes the results as of ADF at level. Because the calculated t values of GDP and

6
Δ3 Yt use for second difference.

11

Electronic copy available at: https://ssrn.com/abstract=1911068

consumption (-0.600268) (-0.361552) respectively are higher than the critical values at all

significant levels. Therefore we do not reject the null hypothesis. Hence it is possible to say that

the series of GDP and consumption are not stationary at first difference and has a unit root.

Table: 3-2 Augmented Dickey Fuller Test at First Difference

Variables t-statistics Level of significance Critical value

1% Critical Value -3.699871

ADF test statistic-GDP -0.600268
5% Critical Value -2.976263

10% Critical Value -2.627420

1% Critical Value -3.699871

ADF test statistic-Cons -0.361552 5% Critical Value -2.976263

10% Critical Value -2.627420

Data about certain indicators of study does not become stationary at level and at first difference,

so we carry on our analysis towards second difference. Table: 3-3 Augmented Dickey Fuller

Test at Second Difference depicts the output of ADF at second difference with intercept. Here

the series become stationary at 1% level of significance as the t-value of GDP is (-8.935281)

and the critical value at 1% significance level is (-3.699871). The t-value is less than the critical

value, it falls at left hand side of the critical value so we can reject null hypothesis and it might

be possible to say that the series of GDP does not have a unit root at second difference.

Table: 3-3 Augmented Dickey Fuller Test at Second Difference

Variables t-statistics Level of significance Critical value

1% Critical Value
-3.699871
ADF test statistic-GDP -8.935281
5% Critical Value -2.976263

10% Critical Value -2.627420

12

Electronic copy available at: https://ssrn.com/abstract=1911068

 1% Critical Value -3.699871

ADF test statistic-Cons -10.25461 5% Critical Value -2.976263

10% Critical Value -2.627420

Calculated value of consumption is (-10.25461) and critical value at 1% significant level is (-

3.699871), so we reject null hypothesis and conclude that the consumption series does not have

a unit root and is stationary at second difference at 1% significant level. Series of data at second

difference were also plotted on graphs, these also does not show any kind of severe trend.

Graph: 3-1 GDP Graph ADF at Second Difference

150

100

50

0 Series1
0 5 10 15 20 25 30
-50

-100

-150

Graph: 3-2 Consumption ADF at Second Difference

13

Electronic copy available at: https://ssrn.com/abstract=1911068

 15

10

0 Series1
0 5 10 15 20 25 30
-5

-10

-15

Conclusion

This study was based on to test the data for stationarity. Testing the data for stationarity is the

first step of data analysis in economics and finance research. Without doing so we cannot apply

the appropriate statistical and econometrics tools to make decisions. We use informal and

formal method to test the data for stationarity. Informal methods encompass of charts and

diagrams while the formal way to test the stationary we use Dickey-Fuller and Augmented

Dickey-Fuller test. We know that Augmented Dickey-Fuller test is commonly used to test the

unit root.

Real data of Gross Domestic Product (GDP) per capita and consumption per capita of India for

the period of 1980-2009 was selected for this purpose. Section one encompasses, scatter plots

and Dickey-Fuller test. Scatter plots portray the increasing trend in both the series. While the

first formal test i.e. Dickey-Fuller test reports that the series are non-stationary at level and at

first difference. But at second difference GDP and consumption become stationary at 1%

significant level. Section two comprises of Augmented Dickey-Fuller test, it does not contain

the different findings as compared to Dickey-Fuller test. We also applied Augmented Dickey-

14

Electronic copy available at: https://ssrn.com/abstract=1911068

Fuller test with time and trend at back end, it also gives almost similar results. There are some

critiques on Augmented Dickey-Fuller test as well, but besides these critiques this test is

considered most important and widely used in research where the time series data is involved.

References

DICKEY D., FULLER W. (1979). – « Distribution of the Estimator for Autoregressive Time
series with a Unit Root », Journal of the American Statistical Association, 74, pp. 427-431.

GRANGER, C.W.J. and NEWBOLD (1974). – « Economic Forecasting: The atheist’s

Viewpoint, in G.A Renton (ed.) », Modeling the economy. London: Heinemann.

APPENDICES

Appendix Table-1
Least Square
Dickey-Fuller Test (GDP)

DF-GLS Test Equation on GLS Detrended Residuals

Dependent Variable: D(GLSRESID)
Method: Least Squares
Sample (adjusted): 1983 2009
Included observations: 27 after adjustments

Variable Coefficient Std. Error t-Statistic Prob.

GLSRESID(-1) 0.003715 0.020181 0.184098 0.8555

D(GLSRESID(-1)) 0.382699 0.196755 1.945057 0.0636
D(GLSRESID(-2)) 0.656065 0.210742 3.113121 0.0047

DF-GLS Test Equation on GLS Detrended Residuals

Dependent Variable: D(GLSRESID)
Method: Least Squares

Variable Coefficient Std. Error t-Statistic Prob.

GLSRESID(-1) -0.105257 0.156294 -0.673457 0.5068

D(GLSRESID(-1)) -0.584431 0.202743 -2.882622 0.0080

15

Electronic copy available at: https://ssrn.com/abstract=1911068

DF-GLS Test Equation on GLS Detrended Residuals
Dependent Variable: D(GLSRESID)
Method: Least Squares

Variable Coefficient Std. Error t-Statistic Prob.

GLSRESID(-1) -1.644567 0.180160 -9.128370 0.0000

Least Square Consumption ( DF)

DF-GLS Test Equation on GLS Detrended Residuals

Dependent Variable: D(GLSRESID)
Method: Least Squares

Variable Coefficient Std. Error t-Statistic Prob.

GLSRESID(-1) 0.001392 0.022161 0.062812 0.9504

D(GLSRESID(-1)) 0.449587 0.171200 2.626096 0.0148
D(GLSRESID(-2)) 0.627739 0.182782 3.434365 0.0022

DF-GLS Test Equation on GLS Detrended Residuals

Dependent Variable: D(GLSRESID)
Method: Least Squares

Variable Coefficient Std. Error t-Statistic Prob.

GLSRESID(-1) -0.059873 0.146361 -0.409076 0.6860

D(GLSRESID(-1)) -0.530184 0.183378 -2.891208 0.0078

DF-GLS Test Equation on GLS Detrended Residuals

Dependent Variable: D(GLSRESID)
Method: Least Squares

Variable Coefficient Std. Error t-Statistic Prob.

GLSRESID(-1) -1.463864 0.171958 -8.512923 0.0000

16

Electronic copy available at: https://ssrn.com/abstract=1911068

Appendix Table-2
Least Square
Augmented Dickey-Fuller

Augmented Dickey-Fuller Test Equation

Dependent Variable: D(GDP)
Method: Least Squares
Sample (adjusted): 1981 2009
Included observations: 29 after adjustments

Variable Coefficient Std. Error t-Statistic Prob.

GDP(-1) 0.089699 0.011732 7.645804 0.0000

C -65.29079 19.05187 -3.427001 0.0020

Augmented Dickey-Fuller Test Equation

Dependent Variable: D(GDP,2)
Method: Least Squares

Variable Coefficient Std. Error t-Statistic Prob.

D(GDP(-1)) -0.093481 0.155733 -0.600268 0.5540

D(GDP(-1),2) -0.602272 0.202178 -2.978925 0.0065
C 15.26094 12.96260 1.177305 0.2506

Augmented Dickey-Fuller Test Equation

Dependent Variable: D(GDP,3)
Method: Least Squares

Variable Coefficient Std. Error t-Statistic Prob.

D(GDP(-1),2) -1.648544 0.184498 -8.935281 0.0000

C 8.954292 7.494656 1.194757 0.2434

Augmented Dickey-Fuller Test Equation

Dependent Variable: D(CONS)
Method: Least Squares

Variable Coefficient Std. Error t-Statistic Prob.

CONS(-1) 0.107370 0.013096 8.198867 0.0000

C -18.63444 3.540483 -5.263248 0.0000

Augmented Dickey-Fuller Test Equation

17

Electronic copy available at: https://ssrn.com/abstract=1911068

Dependent Variable: D(CONS,2)
Method: Least Squares

Variable Coefficient Std. Error t-Statistic Prob.

D(CONS(-1)) -0.051773 0.143197 -0.361552 0.7209

D(CONS(-1),2) -0.563873 0.180742 -3.119770 0.0047
C 1.985283 1.609884 1.233184 0.2294

Augmented Dickey-Fuller Test Equation

Dependent Variable: D(CONS,3)
Method: Least Squares

Variable Coefficient Std. Error t-Statistic Prob.

D(CONS(-1),2) -1.595374 0.155576 -10.25461 0.0000

C 1.540527 1.020295 1.509885 0.1436

Figure: 2-1

18

Electronic copy available at: https://ssrn.com/abstract=1911068

 19

Electronic copy available at: https://ssrn.com/abstract=1911068