Questions on ARDL - How to choose appropriate lags

[Axel Ruben]

Hi fellow STATA-lovers!


I am currently working with an ARDL model, with following variables:

Investments, real interest rate, leading indicator of production, quantitative easing, capital stock.

Investments, of course, are the dependent variable.


Variables are stationary as:
I(0): real interest rate

I(1): investments, leading indicator of production, quantitative easing, capital stock


I am in the process of choosing the lags of each variable, but I am unsure as to which lags to actually choose. Note, the leading indicator can not be lagged since we use it as a lead.

Attached is an image showing the ARDL regression results. Thanks in advance!

Best,
Axel

[Sebastian Kripfganz]

To avoid lagging your leading indicator, which is the second variable in your variable list, add the option lags(. 0 . . .).

Aside from that, 53 observations are insufficient to estimate a model with 6 lags per variable. You hardly have any degrees of freedom left. You need to be much more conservative. Even the default maximum lag order 4 might be too high for meaningful inference.

[Axel Ruben]

Thank you very much Sebastian, and thank you for your prompt response! Would it then be safe for me to proceed with choosing the lag with the lowest P-value as presented in the tables? For example lag3 for QE in the consumption regression? Again, thank you!

[Sebastian Kripfganz]

I am not really sure what you mean by "choosing the lag". The ardl command already selects the optimal lags for you.

[Erick Smith]

Hello Stata lovers,
Dear Sebastian Kripfganz
I am also working with ARDL model.
I'm using four series: tct (exchange rate), bct (current account), cct (capital account), and cft (financial account).
Two of them, tct and bct, are integrated at order 1 (I(1)), while cct and cft are integrated at order 0 (I(0)).
My output of ardl is well, but the Bounds test indicates that there is no cointegration between the variables. Then, I did the second and third step of the estimation, but the bounds test always accepted the null hypothesis of no cointegration relationship.
which of this model I should choose or could you give more suggestions if there is anything else to add?

[Sebastian Kripfganz]

I am afraid there is nothing I can say here. If the bounds test does not reject the null hypothesis, this might by itself be an interesting result.

[Axel Ruben]

Hi again! When trying to do the ARDL command in Stata while including the lagged dependent variable as a regressor, I get an error message saying collinearity. I have tried using the dependent variable lagged with 2 and 3 lags, but neither seems to be working. Any ideas why this might happen and any suggestions on workarounds?

(I'm doing the ARDL with a lagged dependent variable as we faced omitted variable bias in the RESET test)

Many thanks!
Axel

[Sebastian Kripfganz]

Without seeing you command line and Stata output, there is not much we can do to help. You should not specify the lagged dependent variable in the list of independent variables. The first element in the lags() option specifies the number of lags for the dependent variable.

[Axel Ruben]

Beautiful Sebastian, thank you!

When entering the following command: ardl d_ln_investments realinterest d_ln_leadingproduction d_ln_capitalstockest d_ln_qe, maxlags (4) lags(. . 0 . .)
the ARDL gives us Lag1 for the dependent variable. Then, when performing the Ramsey RESET test, I get omitted variables like so:

(changing the lag for the dependent variable to lag2 or lag3 does not work either).

Should we include non-linearity in the equation to get rid of this problem, or is there a better way to proceed with our analysis?

Thanks a million beforehand!

Best,
Axel