2
$\begingroup$

I'm currently looking to run an ARDL model - I'm able to compute results that show cointegration, however there is serial correlation when I run the Durbin-Watson and Breusch-Godfrey tests. To correct for this, I have tried to have larger lags, however this results in most of my coefficients becoming insignificant. I was wondering if there was another way to correct for this, such as using HAC standard errors?

If there is a solution e.g. HAC standard errors, how would I write it into an ARDL model on Stata?

Improve this question
$\endgroup$
8
  • $\begingroup$ You need not worry too much about the statistical significance of explanatory variables when building your model; see "Statistical tests for variable selection" by Rob J. Hyndman. $\endgroup$
    – Richard Hardy
    Commented Mar 30, 2023 at 13:11
  • $\begingroup$ Thank you - does that mean you would recommend increasing lags until serial correlation is eliminated? $\endgroup$
    – woodvalestar
    Commented Mar 30, 2023 at 17:22
  • $\begingroup$ This depends on what you are going to use your model for. $\endgroup$
    – Richard Hardy
    Commented Mar 30, 2023 at 18:15
  • $\begingroup$ Sorry I'm not quite sure what you mean - suppose I am finding out the effect of lagged variables on GDP growth, then is it fine to increase lags (even if this reduces significance) until I remove serial correlation? $\endgroup$
    – woodvalestar
    Commented Mar 31, 2023 at 12:16
  • $\begingroup$ What are you going to use your model for? Prediction? Description? Hypothesis testing? Shmueli "To Explain or to Predict?" (2010) introduces some nuance into the question. $\endgroup$
    – Richard Hardy
    Commented Mar 31, 2023 at 12:21

1 Answer 1

Reset to default
0
$\begingroup$

Testing for statistical significance should not be the guiding principle for building a model; see e.g. see Statistical tests for variable selection by Rob J. Hyndman. Model building comes first, inference comes next. But I think you can implement HAC standard errors directly. If Stata does not support that for ARDL model (perhaps it does?), just formulate your model as a regression where lagged terms are included manually.

Improve this answer
$\endgroup$

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.