Re: st: Re: st: How do I suppress computation of standard errors in stata gmm function?

[Lucas <lucaselastic@gmail.com>]

I'll hop on this good idea and say it'd be nice to turn off estimation
of standard errors in the mixed models, too.  I've had the same
situation, where every iteration is slowed by inverting and testing of
a matrix about which I do not care.  Ideally, one would be able to
turn off the standard errors for some levels while still estimating
them for others.  For example, the standard errors for fixed effects
could be estimated, while the standard errors for the variances could
be switched off.

Such a thing seems possible.  Is it?

Sam

On Fri, Aug 2, 2013 at 6:36 AM, Brian P. Poi <bpoi@stata.com> wrote:
> On 08/01/2013 04:20 AM, Howard Smith wrote:
>>
>> Hi Mike
>>
>>> When you run the estimator once, how much time elapses between the
>>> final iteration and displaying results? That would be an upper bound
>>> on the time savings from eliminating the standard error calculation.
>>
>>
>> I did this upper bound check, and in fact that's why I wanted to turn off
>> the standard errors.
>>
>> Quite a lot of time elapses after the last iteration and the computation
>> of standard errors: several minutes. Even though I supply analytical
>> gradients.
>> I am not sure why it takes so long,
>> perhaps because I have a lot of parameters (50) and a very large dataset.
>>
>> It's a pity they can't be switched off.
>>
>> Howard
>>
>
> Howard,
>
> The ability to switch off -gmm-'s computation of standard errors when doing
> bootstrapping is an interesting idea.  We will look into implementing that.
>
> I'd also like to mention two enhancements to -gmm- that were made available
> in the Stata 13 adofile update of 23jul2013.  These can speed up the
> standard error computations quite substantially.
>
> First, you can now specify equation names within the parameters() option
> when you write a function-evaluator program.  The upshot is that when you
> use this feature, you specify equation-level analytic derivatives rather
> than parameter-level derivatives.  If your moment equations contain linear
> combinations of variables, using equation names makes programming the
> derivatives significantly easier, and, moreover, -gmm- is faster.
>
> Second, if you use numeric derivatives, there is a new option
> 'quickderivatives' that can greatly speed up the computation of the standard
> errors.  In a large majority of cases, using 'quickderivatives' is just as
> accurate as the default computations for the derivatives required for the
> standard errors, yet it is much faster.  In some complicated problems where
> computing the standard errors used to take minutes, specifying
> 'quickderivatives' lets -gmm- get the standard errors in a second or two.
>
> In Stata type
>
> . update query
>
> and follow the instructions to get the latest ado-file update.  Both the
> help file and reference manual entry for -gmm- have been updated to reflect
> these new features.
>
>    -- Brian Poi
>    -- bpoi@stata.com
> *
> *   For searches and help try:
> *   http://www.stata.com/help.cgi?search
> *   http://www.stata.com/support/faqs/resources/statalist-faq/
> *   http://www.ats.ucla.edu/stat/stata/
*
*   For searches and help try:
*   http://www.stata.com/help.cgi?search
*   http://www.stata.com/support/faqs/resources/statalist-faq/
*   http://www.ats.ucla.edu/stat/stata/