Learn how easy it is to sync an existing GitHub or Google Code repo to a SourceForge project! See Demo

Close

#67 optional use of asymptotic approximations

closed
nobody
5
2015-03-12
2011-07-09
Pedro AP
No

It would be nice that there was an option to choose asymptotic aproximations in testing output (t-ratio, test for joint significance, Wald test of linear restriction, etc.) instead of the exact statistics and distributions used to compute p-values under the normality assumption. One might want to use OLS without assuming normality in the same maner that one is not always willing to asume homoscedasticity or lack of serial correlation.

It would be the user choice to prefer believing in normality of errors or in asymptotic approximations in finite samples

Discussion

  • Sven S.
    Sven S.
    2015-03-12

    Hmm, two points here:
    First there are the statistical tables and calculator in gretl which one can use to get these values. But I agree it's manual and time-consuming.

    More importantly, however, if you use asymptotic distributions like N(0,1) instead of t(df), then you use thinner tails. It's not clear why this would be better in a given sample than the status quo. It's not as if the asy approx. is really known to be good in finite samples, we just hope the best. In the end there's no way around the fact that the exact distributions are unknown for non-normality...

    So I'm closing this for now.

    thanks,
    sven

     
  • Sven S.
    Sven S.
    2015-03-12

    • status: open --> closed
    • Group: --> Next_Release_(example)
     
  • Allin Cottrell
    Allin Cottrell
    2015-03-12

    Please note, there is a "set" variable named "robust_z" (in CVS
    and snapshots) which switches to asymptotic p-values for results
    generated with the --robust option.