[Sebastian Kripfganz]

Olav Hose
When you estimate the model in first differences, you should not normally use the ec or ec1 option (which would create differences of differences). You estimate the model in first differences because of lack of evidence of a long-run relationship in your previous step. There is then no long-run relationship in the first-differenced model anymore.

Alex Grisdale
25 observations is almost always not enough for an ARDL estimation, especially with 5 independent variables. There are insufficient degrees of freedom to estimate the model, which is why there are dropped variables. Even if you remove some variables and limit the lag order with option maxlags(), inference with just 25 observations will be very unreliable.

Ronak Parikh
Please see slide 41 of my 2018 London Stata Conference Presentation on how to use Newey-West standard errors with ardl.

[Ronak Parikh]

Dear Sebastian,

Thank you very much for your fast reply - I have a question regarding slide 41: how do I interpret 'cmdline'? If my ARDL input is ardl lngdp lnpatents consumption, lags(4 4 4) ec, is this what I should write in place of 'cmdline'?

[Sebastian Kripfganz]

. webuse lutkepohl2
. quietly ardl ln_consump ln_inc, exog(L(0/3)D.ln_inv) trend(qtr) aic regstore(ardlreg)
. quietly estimates restore ardlreg
. local cmdline `"`e(cmdline)'"'
. gettoken cmd cmdline : cmdline
. newey `cmdline' lag(4)

The local cmdline contains the corresponding command line for the regress command (excluding the command name), not the ardl command. With the ardl option regstore(), the results are stored using the regress command; these are then subsequently recovered with the estimates restore command. The newey command eventually fits the same regression.

To see what is contained in the local cmdline, execute the above code and add the line

. display `"`cmdline'"'

[Olav Hose]

Olav Hose
When you estimate the model in first differences, you should not normally use the ec or ec1 option (which would create differences of differences). You estimate the model in first differences because of lack of evidence of a long-run relationship in your previous step. There is then no long-run relationship in the first-differenced model anymore.

Alex Grisdale
25 observations is almost always not enough for an ARDL estimation, especially with 5 independent variables. There are insufficient degrees of freedom to estimate the model, which is why there are dropped variables. Even if you remove some variables and limit the lag order with option maxlags(), inference with just 25 observations will be very unreliable.

Ronak Parikh
Please see slide 41 of my 2018 London Stata Conference Presentation on how to use Newey-West standard errors with ardl.

Dear Sebastian, perfect and thank you for the reply, that make sense. In cases of no significant value of the F-test, do you know any methods of making it significant, or do I simply need to accept the fact that many of my models lacks cointegration? That would be much appreciated.

[Ronak Parikh]

Dear Sebastian,

Thank you very much - when I execute the following code (with my regression), I receive the following output:

quietly ardl lngdp lnpatents consumption, lags(4 4 4) ec regstore(ardlreg)
quietly estimates restore ardlreg
local cmdline ‘"‘e(cmdline)’"’
gettoken cmd cmdline : cmdline
newey ‘cmdline’ lag(4)
‘cmdline’ invalid name

Would you have any idea how I can fix this?

[Sebastian Kripfganz]

Olav Hose
I cannot recommend any p-hacking strategies. You could of course ask whether the model is correctly specified. There might be important variables missing, the maximum lag order may have to be increased to capture all the dynamics, variable transformations (log) might be needed (if not already done), etc. Once you are confident that the model specification is accurate, you then just have to accept the test result. Often, non-significance can be a meaningful result by itself.

[Sebastian Kripfganz]

Ronak Parikh
You need to use the correct inverted commas. Copy the code from my previous post 423 instead of from the presentation PDF.

[Ronak Parikh]

Dear Sebastian, that's really great - I have been able to obtain estimates. Thank you!

[Alex Grisdale]

Sebastian Kripfganz

Thank you very much for the help, I appreciate it.

[Olav Hose]

Olav Hose
I cannot recommend any p-hacking strategies. You could of course ask whether the model is correctly specified. There might be important variables missing, the maximum lag order may have to be increased to capture all the dynamics, variable transformations (log) might be needed (if not already done), etc. Once you are confident that the model specification is accurate, you then just have to accept the test result. Often, non-significance can be a meaningful result by itself.

Perfect, thank you very much. One final question, I tried out some of the proposals you gave, and when I add extra lags on some of the model, it become statistically significant. However, the lag order of that model is higher as proposed by the BIC criteria which I obtained from the varsoc command. I know it is important not to adjust my model until I get the results which I want, however, would having more or less lags as proposed by the criterium make the model biased?

[Sebastian Kripfganz]

The varsoc command is not the most appropriate one for the single-equation ARDL model; it is more appropriate for VAR models. The ardl command does its own lag selection if you just use the aic or bic option, possibly varying the maximum lag order with the maxlags() option. For the bounds test, it is advisable to use a less parsimonious specification to ensure there is no remaining serial correlation; the AIC is therefore generally preferred over the BIC for this purpose.

[Olav Hose]

The varsoc command is not the most appropriate one for the single-equation ARDL model; it is more appropriate for VAR models. The ardl command does its own lag selection if you just use the aic or bic option, possibly varying the maximum lag order with the maxlags() option. For the bounds test, it is advisable to use a less parsimonious specification to ensure there is no remaining serial correlation; the AIC is therefore generally preferred over the BIC for this purpose.

Perfect, thank you very much. I really appreciate it.

[Alex Grisdale]

Hello once again,

When configuring an ardl in stata and all the variable are stationary at I(1) would you input the command "ardl d.lngdp d.lncpi d.hc, ec btest" for example. So would you have to difference all the variables before putting them in the command?

[Sebastian Kripfganz]

You would not normally specify the variables in first differences. When called with the ec option, the command itself parameterizes the model in first-differenced form (with additional level error correction term).

[Olav Hose]

Hello again,

I do have another question. Everything regarding the ARDL has worked out fine for me, but I do wonder if it possible to illustrade the results with a graph? I am meaning to include how the independent variable of interest affect the dependent variabe, and a graph illustrating the effect first with the initial lags of the short run into the long run? I have not find how to illustrate an ARDL with a graph, so would be very helpfull if anyone had any solution.

Thank you.