Hello,
I'm using R language and I want to apply the ARDL bounds testing in my master thesis. My question is: When I test for stationarity using DF-GLS test can I choose k (the lag length) in order to the variable to be I(0) or I(1)? and after that when I test the model I choose the same lag length as I've chosen in the DF-GLS?
Thank you in advance.
I use information criteria for lag lenght selection (AIC and SIC are the most used in applied econometrics). Lag lenght is not the same if you are applying unit root tests than if you estimate a ARDL bounds test. If you use eviews, you can select the information criteria for lag leght selection. See this blog post by Dave Giles http://davegiles.blogspot.mx/2015/01/ardl-modelling-in-eviews-9.html
I'm using R language but these blogs are useful as they help me to understand the idea.
So for DF-GLS test if the null hypothesis of non-stationary is accepted for k=3 but rejected for k=4 for example i can say that this variable is stationary at levels right?
And also I can't understand for DF-GLS test:
ur.ers(y, type = DF-GLS", model = c("constant", "trend"), lag.max = 4) when to use "constant" and when to use "trend"?
Thank u in advance !
As is indicated, you should use one of those lag length selection criteria of AIC, SC OR HQ. But, make sure that chosen ARDL MODEL should satisfy the underlying assumpons: model should not suffer from autocorrelation, heterescadasticity, coefficients should be stable, no model specification error and errors should have a normal distribution. Good luck!
The lag length selection is sensitive to the maximum lags assigned and, sometimes, the selection criteria, such as AIC and SIC etc. If there is no specification error, it would be fine to adopt rather small number of lags. However, when there is a specification error, more lags would help. In this case, one can drop all insigni cant stationary regressors (lags) to increase degrees of freedom as suggested in Shin, Yu & Greenwood-Nimmo (2014).
google this site http://blog.eviews.com/2017/05/autoregressive-distributed-lag-ardl.html
> So for DF-GLS test if the null hypothesis of non-stationary is accepted >for k=3 but rejected for k=4 for example i can say that this variable is >stationary at levels right?
No. A first assessment of stationarity/non-stationarity should come from theory. There is a good account of the method of selecting the appropriate lag for a DF-GLS test at https://stats.stackexchange.com/questions/151423/which-is-the-best-criterion-for-df-gls-lag-selection.
>And also I can't understand for DF-GLS test:
>ur.ers(y, type = DF-GLS", model = c("constant", "trend"), lag.max = 4) when >to use "constant" and when to use "trend"?
In the traditional ADF unit root test the maintained hypothesis is that the series is non-stationary. There are 3 ADF tests.
The 2 DF-GLS tests are more powerful versions of ADF types 2 and 3 and should be used instead of the corresponding ADF tests
Following. I'm interested in this topic, though not knowledgeable in R i use Eviews mainly
You must take references to the information criteria for lag length selection (AIC and SIC are the most used in applied econometrics). Lag length is not the same if you are applying unit root tests than if you estimate an ARDL bounds test.
- Badges
- Science topic
- Similar topics
- Mathematics