The majority of the blog-related comments and requests for help that I receive come from the one person - called "Anonymous".
(S)he seems to have very broad interests.
Here's a very recent request for help relating to ARDL models - something that I've posted about here and here.
"I am working on income inequality. Can I use ARDL as I have only 27 annual observations? Also does ARDL itself takes care of problem of endogeneity? And what about if there is multicollinearity among explanatory variables - can we still use ARDL? Is any EViews code available to run ARDL?"Taking the questions in order........
- You have nearly 30 years of data, and the thing that matters when testing for unit roots and cointegration is the "span" of the data - not the number of observations. See here. So, in one sense, 30 years is good news in the context of ARDL modelling. On the other hand 30 observations is going to be somewhat limiting when it comes to determining the appropriate lag structure for the ARDL model. I'd be somewhat concerned that you may end up under-stating the lag lengths (specially if more than two variables are involved), and this will have disastrous consequences. So, I'd be very careful!
- ARDL modelling doesn't take care of endogeneity issues.(I assume you're thinking about having regressors that are correlated with the regressors, even asymptotically.) However, this problem is unlikely to arise as long as the errors are serially uncorrelated, because your regressors are typically going to lagged levels, or lagged differences. Also, keep in mind that if cointegration is present, OLS is going to be super-consistent. That is, it's convergence rate will be T-1, as opposed to the usual T-1/2. In short, you should have no problem. However, if this is a concern, you could always use I.V. estimation rather than OLS.
- Don't get me started about multicollinearity! ( See here, here, here, and here.) I doubt if it's a problem - look at all the differencing of the data that you're doing when you estimate an ARDL model.
- To estimate an ARDL model and undertake bounds testing, all you need is a package that will do OLS regression. If you are using just two variables, and you want to automate the lag-length specification for your ARDL using EViews, there's an "add-in" that I discussed some time ago in this post.
Now - I really should get some work done!
Dear Prof. Giles,
ReplyDeleteA question on endogeneity: I'm estimating a standard demand equation with Q as the dependent variable and P as one of the regressors (alongside price of substitute and demand shifter). I'm so far using standard IV (2SLS) with cost shifters as instruments. But, both on economic terms as well as on econometric specification, it makes sense to include lagged Q and lagged P as regressors. Now, as you said, due to endogeneity, try to fit an ARDL model to this demand equation may not be appropriate. My question is: how can I apply IV methods to such dynamic model? Should I instrument the lagged Q and P as well (since Q is a function of lagged Q and lagged P, these two variables are correlated to the contemporaneous error, no?)?
No. Both lagged Q and lagged P are "predetermined" and can serve as their own instruments, provided that the errors are not autocorrelated.
DeleteDear Prof Giles
ReplyDeleteThanks for the valuable information. sir.
Can you tell me sir can we use ARDL model in the situation where Dependent Variable is I(0) and independent Variables are I(1). Thank you sir
Yes, you can. This is clear in the original Pesaran et al. papers.
DeleteHello Sir,
ReplyDeleteThanku Sir for giving a valuable steps for ARDL model.
Can i use ARDL model for Panel Data
Yes, you can.
DeleteHello sir,
ReplyDeletethank you for your job. I have a question: what should I do if the dependent variable is I(2) and the other variables is I(1)?
Thank you
I've noted repeatedly in my posts about ARDL models that they can't be used if any of the variables are I(2).
Delete