Robust standard errors can not be selected for random effect models (it works for FE models, tough) - maybe just the GUI is affected? Or are they simply not implemented for random effect models?
Also implemented is the small sample adjustment in the Stata-way.
My feature request is:
Would you consider to make the small sample adjustment for the variance-covariance optional? It would be nice to have the 'raw' variance-covariance matrix and also, there are various kinds of small sample adjustments, e.g. SAS uses a different one und R's sandwich package supports various weighting schemes.
If you would like to refer to this comment somewhere else in this project, copy and paste the following link:
Also implemented is the small sample adjustment in the Stata-way.
My feature request is: Would you consider to make the small sample
adjustment for the variance-covariance optional? It would be nice to
have the 'raw' variance-covariance matrix and also, there are various
kinds of small sample adjustments, e.g. SAS uses a different one und
R's sandwich package supports various weighting schemes.
My spontaneous reaction would be: If you need to compare stuff at this
level of detail, it's OK to ask the user (you) to apply the correction
factor sqrt((G/(G-1)) * (n-1)/(n-k)) (or its inverse) manually.
Just my 2c,
sven
If you would like to refer to this comment somewhere else in this project, copy and paste the following link:
I am a bit reluctant to have the Stata adjustment as an default silently (just because there are numerous other ways to do this). So maybe we could have a decent documentation on the topic of adjustments and show how to reverse this adjustment (and apply a different one) in ch. 18 of the user manual? I believe, one would need hansl skills to reverse the adjustment, so an example would come in handy too.
If you would like to refer to this comment somewhere else in this project, copy and paste the following link:
I am a bit reluctant to have the Stata adjustment as an default
silently (just because there are numerous other ways to do this). So
maybe we could have a decent documentation on the topic of
adjustments and show how to reverse this adjustment (and apply a
different one) in ch. 18 of the user manual?
Well it doesn't seem to be "silent", because section 17.5 of the guide
is dedicated to it.
However, it's true that section 18.1 appears a little outdated, given
the recent changes:
"In the case of panel data, robust covariance matrix estimators are
available for the pooled and fixed effects model but not currently for
random effects. Please see section 17.4 for details."
There should be a cross-link to section 17.5 instead I guess. The case
of panel RE could be mentioned more explicitly, too. However, as I'm
following this issue just semi-actively, I'm not sure if that would be
completely correct.
And examples are always nice, yes.
cheers,
sven
If you would like to refer to this comment somewhere else in this project, copy and paste the following link:
Thanks for pointing me to section 17.5 where it is indeed mentioned!
Playing the devil's advocat, one could also read that section as if it just mentions possible adjustements. Some of the party of the manual also include general textbook-style comments about statistics, which I value vary much, so this could be one as well.
Maybe would could turn phrases like
"In the least squares case the factor is" into something like
"In the least squares case the factor used by Gretl is".
Yes, cross-reference and an example to reverse the adjustment and/or apply a different adjustment would be very welcomed!
If you would like to refer to this comment somewhere else in this project, copy and paste the following link:
I have just re-read this feature request as well as some of the mailing list messages from April 2016 (this: http://lists.wfu.edu/pipermail/gretl-users/2016-April/011744.html, or this: http://lists.wfu.edu/pipermail/gretl-users/2016-April/011805.html, for example).
The situation is rather confusing. It seems to me that sections 17.4 (Special issues with panel data) and 17.5 (The cluster-robust estimator) of the guide should be merged and/or structured better with respect to the options that gretl offers. Ideally someone who is less confused about these things than I currently am...
A related issue: Gretl still doesn't issue a warning if the user requests "probit ... --random-effects --robust". In the above message Jack explains why the --robust option is ignored, but as I have said before I really don't like this silent fallback. If nobody tells me a good reason, I will soon file this as a new bug ticket. (Perhaps I will try to find the location in the source first and suggest a trivial patch to print the warning, let's see if I can find the right place.)
thanks,
sven
If you would like to refer to this comment somewhere else in this project, copy and paste the following link:
As promised five months ago, this is now officially a bug: a user can specify the impossible combination "probit ... --random-effects --robust" without getting a gretl error or warning; gretl simply ignores the --robust option.
If you would like to refer to this comment somewhere else in this project, copy and paste the following link:
As of today's source version, gretl issues a warning in the "probit ... --random-effects --robust" case. As this was the bug component of this ticket, I'm closing this.
Admittedly some cleanup of the various chapters in the guide would still be helpful...
thanks,
sven
If you would like to refer to this comment somewhere else in this project, copy and paste the following link:
For reference and as a feature request:
this is now implemented in the development version, see e.g.
http://lists.wfu.edu/pipermail/gretl-users/2016-April/011744.html
Also implemented is the small sample adjustment in the Stata-way.
My feature request is:
Would you consider to make the small sample adjustment for the variance-covariance optional? It would be nice to have the 'raw' variance-covariance matrix and also, there are various kinds of small sample adjustments, e.g. SAS uses a different one und R's sandwich package supports various weighting schemes.
Am 07.04.2016 um 14:48 schrieb KTTK:
My spontaneous reaction would be: If you need to compare stuff at this
level of detail, it's OK to ask the user (you) to apply the correction
factor sqrt((G/(G-1)) * (n-1)/(n-k)) (or its inverse) manually.
Just my 2c,
sven
Sven, that is a really good point.
I am a bit reluctant to have the Stata adjustment as an default silently (just because there are numerous other ways to do this). So maybe we could have a decent documentation on the topic of adjustments and show how to reverse this adjustment (and apply a different one) in ch. 18 of the user manual? I believe, one would need hansl skills to reverse the adjustment, so an example would come in handy too.
Am 09.04.2016 um 09:17 schrieb KTTK:
Well it doesn't seem to be "silent", because section 17.5 of the guide
is dedicated to it.
However, it's true that section 18.1 appears a little outdated, given
the recent changes:
"In the case of panel data, robust covariance matrix estimators are
available for the pooled and fixed effects model but not currently for
random effects. Please see section 17.4 for details."
There should be a cross-link to section 17.5 instead I guess. The case
of panel RE could be mentioned more explicitly, too. However, as I'm
following this issue just semi-actively, I'm not sure if that would be
completely correct.
And examples are always nice, yes.
cheers,
sven
Thanks for pointing me to section 17.5 where it is indeed mentioned!
Playing the devil's advocat, one could also read that section as if it just mentions possible adjustements. Some of the party of the manual also include general textbook-style comments about statistics, which I value vary much, so this could be one as well.
Maybe would could turn phrases like
"In the least squares case the factor is" into something like
"In the least squares case the factor used by Gretl is".
Yes, cross-reference and an example to reverse the adjustment and/or apply a different adjustment would be very welcomed!
I have just re-read this feature request as well as some of the mailing list messages from April 2016 (this: http://lists.wfu.edu/pipermail/gretl-users/2016-April/011744.html, or this: http://lists.wfu.edu/pipermail/gretl-users/2016-April/011805.html, for example).
The situation is rather confusing. It seems to me that sections 17.4 (Special issues with panel data) and 17.5 (The cluster-robust estimator) of the guide should be merged and/or structured better with respect to the options that gretl offers. Ideally someone who is less confused about these things than I currently am...
A related issue: Gretl still doesn't issue a warning if the user requests "probit ... --random-effects --robust". In the above message Jack explains why the --robust option is ignored, but as I have said before I really don't like this silent fallback. If nobody tells me a good reason, I will soon file this as a new bug ticket. (Perhaps I will try to find the location in the source first and suggest a trivial patch to print the warning, let's see if I can find the right place.)
thanks,
sven
Ticket moved from /p/gretl/feature-requests/92/
As promised five months ago, this is now officially a bug: a user can specify the impossible combination "probit ... --random-effects --robust" without getting a gretl error or warning; gretl simply ignores the --robust option.
As of today's source version, gretl issues a warning in the "probit ... --random-effects --robust" case. As this was the bug component of this ticket, I'm closing this.
Admittedly some cleanup of the various chapters in the guide would still be helpful...
thanks,
sven