Bookmark and Share

Notice: On March 31, it was announced that Statalist is moving from an email list to a forum. The old list will shut down at the end of May, and its replacement, statalist.org is already up and running.


[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

Re: st: Computation Speed for Bootstrapping


From   Benjamin Volland <volland@econ.mpg.de>
To   statalist@hsphsun2.harvard.edu
Subject   Re: st: Computation Speed for Bootstrapping
Date   Mon, 05 Dec 2011 11:05:36 +0100

Dan,

without knowing your code it is hard to say what's going on. It is particularly puzzling because the only thing that (presumably) changes from your running code to Jeph's suggestion is that you specify the starting values for maximization.

When running the following very simple example:
******************
version 12
set more off
timer clear

webuse grunfeld, clear
egen minvest = mean(invest), by(company)
gen binvest = 1 if invest > minvest
replace binvest = 0 if invest <= minvest
xtset company time

timer on 1
xtprobit binvest kstock mvalue
matrix b = e(b)
set seed 911
xtprobit binvest kstock mvalue, vce(boot, reps(20)) from(b)
timer off 1

timer on 2
set seed 911
xtprobit binvest kstock mvalue, vce(boot, reps(20))
timer off 2

timer list
******************
everything works fine on my machine. Moreover, -timer- reveals that Jeph's suggestion substantially increases estimation speed in this simple case.
. timer list
   1:      3.50 /        1 =       3.5000
   2:      4.16 /        1 =       4.1560

Another thought on the speed of your estimation: it may be an idea to check your independents for (multi)collinearity, as this may substantially slow down convergence.

Best, Ben


On 04/12/2011 05:05, Danny Dan wrote:
Thank you so much Ben for replying to my query. I have tried
implementing your's and Jeph's suggestions but now it is giving me the
following error r(2000):

insufficient observations to compute bootstrap standard errors
no results will be saved

However, without those additional commands my bootstrap is running,
although the computation is extremely slow (taking almost 4 days for
only 50 reps. My data set contains 10,000 obs.).

Also do you or any statalist users have any idea why the problem
r(2000) (insufficient observations to compute bootstrap standard
errors, no results will be saved) occurs and how to fix it? If yes,
then please let me know.

Thanks and Best wishes.

Dan



On Fri, Dec 2, 2011 at 5:08 AM, Benjamin Volland<volland@econ.mpg.de>  wrote:
Hi Danny,

your most easy option is to follow Jeph's second suggestion.
That is, run your estimation once on the full sample. Then use
- matrix b = e(b) - and then simply re-use the same matrix as starting
values for each bootstrap iteration, using the -from(b)- option in your
-xtprobit- command.
You'll probably have to check whether this increases estimation speed
substantially using a small number of bootstrap interations, and simply
compare the time with and without setting initial values.
If you want to use different starting values in each iteration (though I
somehow fail to see the improvement given that each bootstrap iteration
simply draws a new sample from the original data), i.e.
write your own bootstrap program, check the Sata FAQ from the UCLA on how to
do it:
http://www.ats.ucla.edu/stat/stata/faq/ownboot.htm
Again, whether that substantially increases the speed of your estimation you
would have to check.

Best, Ben



On 01/12/2011 23:41, Danny Dan wrote:

Thank you so much for the answer Jeph. However, I do not know how to
write the code as mentioned in your reply. If you can give me further
insight that would be really helpful.

Thanks once again.

On Sun, Nov 27, 2011 at 10:06 AM, Jeph Herrin<stata@spandrel.net>    wrote:

One trick is to re-use the estimation from each model as the starting
values
for the next one. So after calling xtprobit, save the matrix e(b) and
then
use it with the -from(b)- option.

   xtprobit ...
   matrix b=e(b)
   xtprobit ...., from(b)

But to do this, I think you'll have to write your own bootstrap code (not
too difficult).
Or you may get by with using -bootstrap:-, if you reuse the same matrix
for
each
bootstrap.

Hope this helps,
Jeph



On 11/26/2011 10:31 PM, Danny Dan wrote:


Dear All,

How to increase computation speed while generating standard errors
using bootstrapping for panel data probit model (xtprobit,
vce(bootstrap))?

Reduction in the number of replications (reps()) is of no help.

Please let me know.

Thank you,

Regards,

Malini
*
*   For searches and help try:
*   http://www.stata.com/help.cgi?search
*   http://www.stata.com/support/statalist/faq
*   http://www.ats.ucla.edu/stat/stata/


-----
No virus found in this message.
Checked by AVG - www.avg.com
Version: 2012.0.1873 / Virus Database: 2101/4642 - Release Date:
11/27/11



*
*   For searches and help try:
*   http://www.stata.com/help.cgi?search
*   http://www.stata.com/support/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/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/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/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/statalist/faq
*   http://www.ats.ucla.edu/stat/stata/


© Copyright 1996–2014 StataCorp LP   |   Terms of use   |   Privacy   |   Contact us   |   Site index