ECS4863

SUGGESTED SOLUTIONS TO ACTIVITY 1.2

1.7 Activity 1.2: Testing for Cointegration: Engle-Granger (EG) Test

This activity deals with the following assessment criteria:

- Demonstrate tests for unit roots
- Construct tests for cointegration between long-run variables

The aim of this activity is to estimate a government revenue function for South Africa. The
theoretical specification of government revenue (REV) may be specified as:

REV = f (M3, YD, i)

Where:

- M3 (money supply or wealth) is expected to be positively correlated

- YD (disposable income) is expected to be positively correlated

- i (interest rate) is expected to be negatively correlated

The data is available on myUnisa/additional resources/Activity data: Activity 1.2 data

1. Generate the necessary real values of the data and log all variables.

Use the genr (generate) function in EViews and create the following variables:

RREV=rev/cpi*100

RYD=nyd/CPI*100

rm3=m3/cpi*100

infl=(cpi/cpi(-1)-1)*100

rprime=prime-infl

Open Rubric
2. Test the variables for their order of integration (remember that we already tested M3
and YD in activity 1.1)
Series Model ADF lags ADF

ττ τ µ τ

LRREV Trend and intercept ( τ τ ) 1 -3.095

Intercept ( τ µ ) 0 0.257

None ( τ ) 0 2.881

D(LRREV) Trend and intercept ( τ τ ) 0 -3.487*

Intercept ( τ µ ) 0 -3.675**

None ( τ ) 0 -2.660**

Lrprime Trend and intercept ( τ τ ) 0 -1.868

Intercept ( τ µ ) 0 -0.617

None ( τ ) 0 -0.613

D(Lrprime) Trend and intercept ( τ τ ) 3 -3.808**

Intercept ( τ µ ) 0 -4.135***

None ( τ ) 0 -4.195***

* Statistically significant at the 10% level

** Statistically significant at the 5% level
*** Statistically significant at the 1% level

The results of the unit root test indicate that both LRREV and LRPRIME becomes stationary
after differencing once. This indicates that they are I(1) variables. You may assume (as
proven in activity 1.1) that RM3 and RYD are also both I(1).

[Interesting to note: if you re-do the tests for RM3 and RYD you will see that there are
differences in values between them and the ones you did for the same variables in Activity
1.1. The general conclusion (i.e. whether the variables are I(0), I(1), etc.) should however
remain the same. The reason for the difference is of course the different time periods (years)
and number of observations used. Choosing which time period, and hence which number of
observations to work with, are often difficult decisions the econometrician needs to take.
Also watch out for structural breaks in the data.]
3. Estimate the following potential long-run cointegration equation, and completed the
table below:

Log(RREV) = β 0 + β1 Log(RYD) + β 2 Log(RM3)

Estimating an equation in EViews:

Press and hold Ctrl (control) on your keyboard

Select the variables, always starting with your dependent variable (in this
case LRREV) and then the rest.
[As long as Ctrl is pressed you can select/deselect variables as you like. They
will be highlighted in blue or not as you select/de-select them.]

Once a variable has been selected, it will be indicated/highlighted by a blue

block around the variable.

Let go of Ctrl.

Making sure that your cursor is in any of the blue blocks, right-click and
select “Open/As Equation” …

An Equation Estimation pop-up will appear:

In this pop-up you can manually type in the variables or change them around.
For example, if you prefer the constant term to appear first (in front of the
other explanatory variables), you can delete it (backspace) and type in front:
(i.e. lrrev c lryd lrm3). Remember to keep spaces between the variables.
Check the Estimation settings, that the estimation method is Least Squares
(default option) and that the sample range (in this case 1990 to 2013) is
correct (this will be included by EViews as default).
Click “OK”.

Dependent variable: LRREV

Variable Coefficient

LRYD 0.34

LRM3 0.51

C 0.75

How do we evaluate the (potential!) cointegration (or long-run)

equation?

• Economic theory rules (especially when you are dealing with small
sample sizes). What we want is for the coefficients to have the correct
signs and magnitudes. We do “look” at the t-stats for safety (i.e. to
make sure there is at least some statistical significance), however,
technically statistical inference is invalid due to the times series (likely)
being non-stationary.
• Make sure to estimate the equation in levels (do not difference any of
the variables at this stage. Why?).
• No lags.
• Use dummy variables sparingly (only if you have very good reason to
suspect a structural break in the data).
• With or without a constant.
• The other statistics (e.g. R², F, t-stats) are invalid, so do not report or
evaluate them at this stage.
• NB: Look for the minimum number of explanatory variables that result
in stationary residuals (this will be tested in the next step / Question
4) and that is consistent with economic theory.
Using the above guidelines to evaluate our (potential) cointegration equation, it is
evident that both LRYD and LRM3 are positively related to LRREV (i.e. they both
have a positive sign), which is correct according to our a priori (economic theory)
expectations. The magnitude of the coefficients also seems relevant – in this
case they are both between 0 and 1, and if LRYD increases by 1% we expect
LRREV to increase by 0.34%, while if LRM3 increase by 1% we expect LRREV
to increase by 0.51%.

(Determining the magnitude is much more difficult as various samples and time
periods could lead to different magnitudes. This is where the study of economic
theory or academic papers prior to doing econometric modelling becomes very
important, so that the researcher has a good grasp of the theory and also can be
able to set rational expectations.)

This is also a good time to save your (potential) cointegrating equation. You do
this by selecting “Name”, and then typing any name you would like to give the
long-run equation. Suggestions could be eq01 (default), eq_lr (for long run),
eq_coint (for cointegration), etc. Just remember to either type one word, or the
underscore (_) sign to connect terms. Click “OK”. You will see a “new variable”
in the workfile, with the same name as the one you gave your equation. Also,
there will be an equal to / = sign next to the variables, indicating that it is an
equation you saved. You can now close the equation, and every time you select
(double-click) the new variable, EViews will open up the equation automatically
(thus saving you time not to have to reselect each time).

Also make sure to save you work file itself regularly. To do this, you go to the top-
left corner of EViews interface, and then select “File/Save as” … (similar as how
you would save for example any MS Word or Excel file). You will see that the file
is saved as a .wf1 file, indicating that it is an EViews work file.

4. Test for cointegration between the variables. Complete the table.

(This can be done in two ways, i.e. either to make a repressor group or to
generate the residual and then test for unit root. We preferred that you use the
latter, i.e. generate the residual and then test for unit root. The reason is that if
the residuals test stationary, you will use the new series in the rest of the
modelling, in which case the series has already been generated. This will
likely save some time and possible confusion. Students who would also like to
use the repressor group method can contact the lecturers for guidance.)

Once the researcher is satisfied that the signs and magnitudes of the coefficients
are consistent with economic theory, he/she can proceed to formally test if the
residuals of the (potential) cointegrating equation are indeed stationary.
The hypothesis of the cointegration test is as follows:

H0: No cointegration.

How to generate the residuals:

Open the (potential) cointegration equation you saved in the previous step.

Click “Proc/Make residual series” …

“A Make Residuals” pop-up will appear. For residual type select “Ordinary”
(default option). Also, this is a good time to name your residual series. The
default is resid01, but you should type in any new name you want for the new
series; suggestions: res_rev (residual for the revenue equation) or resid_rev, etc.
If you do not type in a new name, EViews will generate a new variable called
resid01.

You can also change the name later by selecting the resid01 variable and
renaming it, by repeating the steps by clicking on “Name”/etc … . Keep in mind
that this residual series will feature very prominently in the rest of the model
process, so just make sure that you select a name that you will be able to identify
easily later in the process (all the equations can become confusing!)

Click “OK”.

The new series/variable will appear in the workfile.

Open (double click) the new variable (res_rev).

Click View/Unit root “test” …

Select:

Test for unit root in: Level

Include in test equation: None

Leave the rest as default (i.e. Augmented Dickey-Fuller test and Automatic Lag
length selection, using the Schwartz Info Criterion).

Click “Ok”..
 Model Lags τ p-value

LRREV 3 -4.395 0.0002***

The results indicate that the variable is statistically significant (at 1% level, i.e.
three stars, which means that we can reject the null hypothesis (of no
cointegration).

5. Can we conclude that the variables in the long-run equation are

cointegrated?

Yes, as proven in Question 4.

IF AND ONLY IF YOUR RESULTS INDICATE THAT THE EQUATION IS

INDEED COINTEGRATED, may you proceed with the ECM. IF NOT, TRY AND
TRY AGAIN!

(Unfortunately this is one of the major frustrations often faced by econometrists,

i.e. to get cointegrating results. This is also where the notion comes from that
econometric modelling is often half science, half art. But try and keep your
potential equation as simple as possible and it is likely that after a few re-runs/
adjustments you will be rewarded. And remember that economic theory is king!)

6. Evaluate the signs and magnitudes of the variables in the long-run

equation. Do they make economic sense?

Both LRYD and LRM3 are positively related to LRREV (i.e. they both have a
positive sign), which is correct according to our a priori (economic theory)
expectations. The magnitude of the coefficients also seems relevant – in this
case they are both between 0 and1, and if LRYD increases by 1% we expect
LRREV to increase by 0.34%, while if LRM3 increase by 1% we expect LRREV
to increase by 0.51%.

7. Why do we not report any other statistical inference (e.g. t-values, F-

values etc) for the long-run equation?

They are invalid. See the explanation in Question 3 above.