ADF test

[glu]

As far as I understand, Eviews selects the lag length in the ADF test by minimizing a prescribed criterion (AIC, or SIC, or Hannah-Quinn). However, testing an 84-observation time series (attached), I find that selected lag lengths do not correspond to the minimums; and selected values of the criteria also are not the minimal ones (see the table below). Could you please explain the algorithm of the lag length selection behind this?

Lag
length AIC SIC Hannan-Quinn
0 -2.558672 -2.529529* -2.546964
1 -2.539901 -2.481200 -2.516333
2 -2.526917 -2.438233 -2.491336
3 -2.571450 -2.452349 -2.523699
4 -2.724223 -2.574258 -2.664143
5 -2.927630 -2.746345 -2.855058*
6 -2.943586 -2.730513 -2.858359
7 -3.095566* -2.850226** -2.997516**
8 -3.065688 -2.787590 -2.954647
9 -3.058219 -2.746859 -2.934014
10 -3.091807 -2.746669 -2.954263
11 -3.098775** -2.719331 -2.947717
* Selected by EViews
** True minimum

[EViews Gareth]

You've probably calculated the manual AIC values incorrectly. As with all lag-length selection routines, you have to be careful to use the same sample for each test.

I quickly did a manual calculation of the AIC for 7 lags, and got -3.158232, which doesn't match with your -3.095566.

[glu]

All outputs are those from EViews (see the file attached to my first message). Besides, I checked the values of the criteria for different lag lengths in a different software (Matrixer) and obtained the same values.

[EViews Gareth]

I think you're getting confused between the final ADF equation output, and the equations that are used for performing lag-length selection. As I said above, when calculating the lag-length, the same sample has to be used for all lag lengths. Once an optimal lag-length has been chosen, the sample is expanded for the final equation. Obviously this can mean that the equation used for the selection process can be different (use a different number of observations) from the one used for the final ADF.

[glu]

Thank you!
The problem is in that the EViews guides do not clearly report the algorithm of lag-length selection. Ng and Perron's paper (Oxford Bull. of Ec. & Stat., 2005, 67(1) ) is misleading, referring to EViews as applying what the call "Method 2", while in fact it applies "Method 1". (They refer to the 1997 version of EViews, though.)

In addition: the @hq does not work in Eviews 6; its Help reports no results regarding "@hq" and even "Hannan-Quinn".

[EViews Gareth]

Code: Select all

create u 100 series y=nrnd series x=nrnd equation e1.ls y c x scalar hq = e1.@hq

Seems to work for me.

In EViews 6, you're better off using the PDFs (assuming your PDF reader has a search feature) than the help system.

[alexyflemming]

Dear Eviews Gareth,

I understand what you say theoretically. However, I cannot find your -3.158232 value for lag 7. Hıw did you find -3.158232 in Eviews. I am using EViews 7.2.
ADF(maxlag11, AIC)-->gives optimal lag 7 with 76 observations.


I tried to find Glu's value. Here are my results:
ADF; Level.
Automatic/ADFeq Optimal AIC SIC HQ
----------- ---------- ----- ---- ------
0,none 0 Glu's values for lag 0.
1,none 0 Glu's values for Lag 0. I don't know how Glu found values for Lag 1.
2,none 0 Glu's values for Lag 0. I don't know how Glu found values for Lag 2.
3,none 3 Glu's values for lag 3.
4,none 1 Glu's values for lag 1.
5,none 5 Glu's values for lag 5.
6,none 6 Glu's values for lag 6.
7,none 6 Glu's values for lag 6. I obtained Glu's Lag 7 values in Lag 11 values.
8,none 6
9,none 6
10,none 6
11,none 7 Glu's values for lag 7.
12,none 7
13,none 7
14,none 7
15,none 11
16,none 11
17,none 7
18,none 11
19,none 5
28,none 10 Glu's values for lag 10.
36,none 0
40,none 39
41,none EVIEWS REFUSED TO CALCULATE!

Could you please explain, how to expand the sample for the final ADF output? Which menus should we use?
For lag 7, assume we used the same sample as with those of lags 1,2,...,6,8,9,10,11. How many obs should we use for final ADF test after series expansion (76, 84, ..observations)? And most importantly HOW? THANK YOU VERY MUCH IN ADVANCE.

[EViews Gareth]

Simply set the workfile sample, by using the smpl command, or double clicking on the "Sample: " part of the workfile prior to running the ADF test. With a maximum of 11 lags, in Glu's case, the sample should be 1995m1 2000m12.

Running an equation of:

Code: Select all

d(ch07) ch07(-1) d(ch07(-1)) d(ch07(-2)) d(ch07(-3)) d(ch07(-4)) d(ch07(-5)) d(ch07(-6)) d(ch07(-7))

with that sample produces the required AIC value.

[alexyflemming]

Very very very thanks, Eviews Gareth.

I got it and I obtained the result you gave. Also, to clarify further for the others:

In ADF regressions, since there is DELTAyt on left hand side of the regression equation, the number of observations used is: For lag 0: T - 1 (T:= number of observations).

Just for the same token; For lag 1: T - 2; for lag 2: T - 3; ...; For lag k: T - (k+1).

Hence, in the above Glu's example, there is 11 lags; i.e., the number of obs that must be used in ADF regressions is 84 - (11+1) = 84 - 12 = 72. I.e., one must exclude the 12m = 1year obs from the sample. That's why, sample turns to be: 1995M1-2000M12.

Best and Warm Regards,

Alexy Flemming