Quantcast

Suest v/s biprob in stata 11

classic Classic list List threaded Threaded
8 messages Options
Reply | Threaded
Open this post in threaded view
|  
Report Content as Inappropriate

Suest v/s biprob in stata 11

Prakash Kashwan
Dear All,
I have a couple of queries regarding how suest works in Stata 11. The
new user friendly menus use biprob command, which gives me different
results compared to when I use suest posestimation command - I wonder
if biprob does utilize the leverage offered by correlating residuals
of two models involved. In any case, I get different results with
these two commands. Moreover, when I run suest it doesn't give me any
test statistics so that I know if using suest (as opposed to
independent robust estimates) gives me additional leverage. I have
extensively searched for clues on these issues but no success at all.
Am I missing something obvious here? I will appreciate help with
figuring out how to make test give me test statistics following the
suest command.
Prakash
*
*   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/
Reply | Threaded
Open this post in threaded view
|  
Report Content as Inappropriate

AW: Suest v/s biprob in stata 11

Martin Weiss-5

<>


" The
new user friendly menus use biprob command, which gives me different
results compared to when I use suest posestimation command"


What does "the new user friendly menus" mean? And what is the -biprob-
command/where did you obtain it???



HTH
Martin


-----Ursprüngliche Nachricht-----
Von: [hidden email]
[mailto:[hidden email]] Im Auftrag von Prakash Kashwan
Gesendet: Sonntag, 4. Juli 2010 19:22
An: [hidden email]
Betreff: st: Suest v/s biprob in stata 11

Dear All,
I have a couple of queries regarding how suest works in Stata 11. The
new user friendly menus use biprob command, which gives me different
results compared to when I use suest posestimation command - I wonder
if biprob does utilize the leverage offered by correlating residuals
of two models involved. In any case, I get different results with
these two commands. Moreover, when I run suest it doesn't give me any
test statistics so that I know if using suest (as opposed to
independent robust estimates) gives me additional leverage. I have
extensively searched for clues on these issues but no success at all.
Am I missing something obvious here? I will appreciate help with
figuring out how to make test give me test statistics following the
suest command.
Prakash
*
*   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/
Reply | Threaded
Open this post in threaded view
|  
Report Content as Inappropriate

Re: AW: Suest v/s biprob in stata 11

Prakash Kashwan
In reply to this post by Prakash Kashwan
Hi Martin,

Thanks so much for a prompt response. By user friendly menus, I meant
the SPSS kind of interface that Stata 11 has to offer. In the process
of responding to your query I have also discovered that Stata 11 does
offer a similar menu for the suest command, but still doesn't give me
test statistics. Below, I describe the two different ways in which
Stata 11 carries out 'Seemingly Unrelated Estimation'

Menu -1:  Statistics > Statistics > Binary Outcomes > Seemingly
unrelated estimation

This menu runs the following command

 biprobit (probit model 1) (probit model2), vce(robust) noskip
difficult nonrtolerance


Menu -2: Statistics > Postestimation > Tests > Seemingly unrelated estimation

This one runs the suest command

suest probitModel1 probitModel2, vce(robust)

And, this is the one I want to use, but it doesn't give me test
statistics; basically to find out if suest offers me additional
leverage (as compared to each of the models run independently).


Many thanks,
Prakash


On Sun, Jul 4, 2010 at 3:20 PM, Martin Weiss <[hidden email]> wrote:

>
> <>
>
>
> " The
> new user friendly menus use biprob command, which gives me different
> results compared to when I use suest posestimation command"
>
>
> What does "the new user friendly menus" mean? And what is the -biprob-
> command/where did you obtain it???
>
>
>
> HTH
> Martin
>
>
> -----Ursprüngliche Nachricht-----
> Von: [hidden email]
> [mailto:[hidden email]] Im Auftrag von Prakash Kashwan
> Gesendet: Sonntag, 4. Juli 2010 19:22
> An: [hidden email]
> Betreff: st: Suest v/s biprob in stata 11
>
> Dear All,
> I have a couple of queries regarding how suest works in Stata 11. The
> new user friendly menus use biprob command, which gives me different
> results compared to when I use suest posestimation command - I wonder
> if biprob does utilize the leverage offered by correlating residuals
> of two models involved. In any case, I get different results with
> these two commands. Moreover, when I run suest it doesn't give me any
> test statistics so that I know if using suest (as opposed to
> independent robust estimates) gives me additional leverage. I have
> extensively searched for clues on these issues but no success at all.
> Am I missing something obvious here? I will appreciate help with
> figuring out how to make test give me test statistics following the
> suest command.
> Prakash
> *
> *   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/
>



--
--
Prakash Kashwan <http://mypage.iu.edu/~pkashwan/>
School of Public & Environmental Affairs (SPEA)
Workshop in Political Theory and Policy Analysis; Indiana University,
Bloomington
Research Fellow (2009) - International Foundation for Science (Sweden)

*
*   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/
Reply | Threaded
Open this post in threaded view
|  
Report Content as Inappropriate

Re: Suest v/s biprob in stata 11

Maarten buis
In reply to this post by Prakash Kashwan
--- On Sun, 4/7/10, Prakash Kashwan wrote:
> I have a couple of queries regarding how suest works in
> Stata 11. The new user friendly menus use biprob command,
> which gives me different results compared to when I use
> suest posestimation command - I wonder if biprob does
> utilize the leverage offered by correlating residuals
> of two models involved.

Yes, you can see in the example below that when you
constrain that correlation to 0, both -biprobit- and
-suest- give the same estimates.

*----------------------- begin example ------------------
webuse school
biprobit (private = logptax loginc years)     ///
         (vote = logptax years)
est store biprobit

probit private logptax loginc years
est store a

probit vote logptax years

est store b
suest a b
est store suest

constraint 1 [athrho]_cons=0
biprobit (private = logptax loginc years)      ///
         (vote = logptax years), constraint(1)
est store biprobit_constrained

est table biprobit suest biprobit_constrained, ///
    equations(1, 2)
*-------------------- end example -----------------
(For more on examples I sent to the Statalist see:
http://www.maartenbuis.nl/example_faq )

Note that -biprobit-, and many other Stata commands,
maximizes the likelihood with respect to the Fisher's
z transformation (or the inverse hyperbolic tangent
of the correlation). This transformed correlation
is stored in the parameter _cons in the equation
athrho. The correlation is 0 when this transformed
correlation is 0, so the appropriate constraint is
-constraint 1 [athrho]_cons=0-.

Hope this helps,
Maarten

--------------------------
Maarten L. Buis
Institut fuer Soziologie
Universitaet Tuebingen
Wilhelmstrasse 36
72074 Tuebingen
Germany

http://www.maartenbuis.nl
--------------------------


     

*
*   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/
Reply | Threaded
Open this post in threaded view
|  
Report Content as Inappropriate

Re: Suest v/s biprob in stata 11

Prakash Kashwan
Hello Maarten (and apologies for misspelling your name in my previous email),
Thanks much for this explanation. From your explanation it seems as if
biprob accounts for the correlation between the residuals from
constituent models while suest does not (only when that is true, a
constrained biprob will give the same result as a suest, which has not
been constrained in the example you cite). Am I interpreting it well?
If this is indeed the case, it is problematic because the way stata
runs it, suest is a postestimation command, which is supposed to look
for correlation between the residuals. Am I to assume that suest is
not doing what it is supposed to do, and I should use biprob instead
of using suest?  What purpose does suest (as post-estimation command)
serve then?
I will appreciate these clarification.
Best Regards,
Prakash

On Mon, Jul 5, 2010 at 8:02 AM, Maarten buis <[hidden email]> wrote:

> --- On Sun, 4/7/10, Prakash Kashwan wrote:
>> I have a couple of queries regarding how suest works in
>> Stata 11. The new user friendly menus use biprob command,
>> which gives me different results compared to when I use
>> suest posestimation command - I wonder if biprob does
>> utilize the leverage offered by correlating residuals
>> of two models involved.
>
> Yes, you can see in the example below that when you
> constrain that correlation to 0, both -biprobit- and
> -suest- give the same estimates.
>
> *----------------------- begin example ------------------
> webuse school
> biprobit (private = logptax loginc years)     ///
>         (vote = logptax years)
> est store biprobit
>
> probit private logptax loginc years
> est store a
>
> probit vote logptax years
>
> est store b
> suest a b
> est store suest
>
> constraint 1 [athrho]_cons=0
> biprobit (private = logptax loginc years)      ///
>         (vote = logptax years), constraint(1)
> est store biprobit_constrained
>
> est table biprobit suest biprobit_constrained, ///
>    equations(1, 2)
> *-------------------- end example -----------------
> (For more on examples I sent to the Statalist see:
> http://www.maartenbuis.nl/example_faq )
>
> Note that -biprobit-, and many other Stata commands,
> maximizes the likelihood with respect to the Fisher's
> z transformation (or the inverse hyperbolic tangent
> of the correlation). This transformed correlation
> is stored in the parameter _cons in the equation
> athrho. The correlation is 0 when this transformed
> correlation is 0, so the appropriate constraint is
> -constraint 1 [athrho]_cons=0-.
>
> Hope this helps,
> Maarten
>
> --------------------------
> Maarten L. Buis
> Institut fuer Soziologie
> Universitaet Tuebingen
> Wilhelmstrasse 36
> 72074 Tuebingen
> Germany
>
> http://www.maartenbuis.nl
> --------------------------
>
>
>
>
> *
> *   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/
>



--
--
Prakash Kashwan <http://mypage.iu.edu/~pkashwan/>
School of Public & Environmental Affairs (SPEA)
Workshop in Political Theory and Policy Analysis; Indiana University,
Bloomington
Research Fellow (2009) - International Foundation for Science (Sweden)

*
*   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/
Reply | Threaded
Open this post in threaded view
|  
Report Content as Inappropriate

Re: Suest v/s biprob in stata 11

Maarten buis
--- On Mon, 5/7/10, Prakash Kashwan wrote:
> From your explanation it seems as if biprob accounts for
> the correlation between the residuals from constituent
> models while suest does not. Am I interpreting it well?

Yes

> If this is indeed the case, it is problematic because the
> way stata runs it, suest is a postestimation command,
> which is supposed to look for correlation between the
> residuals. Am I to assume that suest is not doing what it
> is supposed to do, and I should use biprob instead of
> using suest?  What purpose does suest (as post-estimation
> command) serve then?

The way I understand this is that the inference takes this
correlation into account in a way that is similar to the
way -robust- standard errors take heteroskedasticity into
acount without estimating the changing error variance.
All that is necesary is that your model for the mean(s) is/are
correct. You can see that in the simulation below. You
want the p-values of both test to be uniformly distributed
as they test an hypothesis that is true in the population,
as is discussed in this post:
<http://www.stata.com/statalist/archive/2010-06/msg01191.html>

*----------------- begin example -----------------
matrix C = (1, .25 \ .25, 1)

program drop _all
program sim, rclass
        drop _all

        drawnorm e1 e2, n(1000) corr(C)
        gen x = rnormal()

        gen y1=  ( x + e1) > 0
        gen y2 = (-x + e2) > 0

        probit y1 x
        est store a
        probit y2 x
        est store b

        suest a b
        test [a_y1]x = - [b_y2]x
        return scalar p_suest = r(p)
       
        biprobit y1 y2 = x
        test [y1]x = -[y2]x
        return scalar p_biprobit = r(p)
end

set seed 12345
simulate p_biprobit = r(p_biprobit) ///
         p_suest    = r(p_suest),   ///
         reps(10000) : sim

hangroot p_suest,                   ///
         dist(uniform) par(0 1)     ///
         susp notheor ci
                 
hangroot p_biprobit,                ///
         dist(uniform) par(0 1)     ///
         susp notheor ci
*------------------ end example ----------------
(For more on examples I sent to the Statalist see:
http://www.maartenbuis.nl/example_faq )

This example requires -hangroot-, which you can download
by typing in Stata -ssc install hangroot-.

Hope this helps,
Maarten

--------------------------
Maarten L. Buis
Institut fuer Soziologie
Universitaet Tuebingen
Wilhelmstrasse 36
72074 Tuebingen
Germany

http://www.maartenbuis.nl
--------------------------


     

*
*   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/
Reply | Threaded
Open this post in threaded view
|  
Report Content as Inappropriate

Re: Suest v/s biprob in stata 11

Prakash Kashwan
Thanks so much, Maarten,
I think you may have understood my problem differently than I did.


"All that is necesary is that your model for the mean(s) is/are
correct. ......... You want the p-values of both test to be uniformly
distributed
as they test an hypothesis that is true in the population"

I am conducting the two tests on same set of observations (so, the
population is not divided into two groups). Moreover, in my case, I am
using the same set of independent variables to estimate probabilities
for two different dependent variables, and after I have run the
'suest' command, I would like to figure out if the suest, did find
significant correlation between the two error matrices, and if it
makes sense for me to use suest in this case.

Many Thanks again in advance,
Prakash

On Mon, Jul 5, 2010 at 9:56 AM, Maarten buis <[hidden email]> wrote:

> --- On Mon, 5/7/10, Prakash Kashwan wrote:
>> From your explanation it seems as if biprob accounts for
>> the correlation between the residuals from constituent
>> models while suest does not. Am I interpreting it well?
>
> Yes
>
>> If this is indeed the case, it is problematic because the
>> way stata runs it, suest is a postestimation command,
>> which is supposed to look for correlation between the
>> residuals. Am I to assume that suest is not doing what it
>> is supposed to do, and I should use biprob instead of
>> using suest?  What purpose does suest (as post-estimation
>> command) serve then?
>
> The way I understand this is that the inference takes this
> correlation into account in a way that is similar to the
> way -robust- standard errors take heteroskedasticity into
> acount without estimating the changing error variance.
> All that is necesary is that your model for the mean(s) is/are
> correct. You can see that in the simulation below. You
> want the p-values of both test to be uniformly distributed
> as they test an hypothesis that is true in the population,
> as is discussed in this post:
> <http://www.stata.com/statalist/archive/2010-06/msg01191.html>
>
> *----------------- begin example -----------------
> matrix C = (1, .25 \ .25, 1)
>
> program drop _all
> program sim, rclass
>        drop _all
>
>        drawnorm e1 e2, n(1000) corr(C)
>        gen x = rnormal()
>
>        gen y1=  ( x + e1) > 0
>        gen y2 = (-x + e2) > 0
>
>        probit y1 x
>        est store a
>        probit y2 x
>        est store b
>
>        suest a b
>        test [a_y1]x = - [b_y2]x
>        return scalar p_suest = r(p)
>
>        biprobit y1 y2 = x
>        test [y1]x = -[y2]x
>        return scalar p_biprobit = r(p)
> end
>
> set seed 12345
> simulate p_biprobit = r(p_biprobit) ///
>         p_suest    = r(p_suest),   ///
>         reps(10000) : sim
>
> hangroot p_suest,                   ///
>         dist(uniform) par(0 1)     ///
>         susp notheor ci
>
> hangroot p_biprobit,                ///
>         dist(uniform) par(0 1)     ///
>         susp notheor ci
> *------------------ end example ----------------
> (For more on examples I sent to the Statalist see:
> http://www.maartenbuis.nl/example_faq )
>
> This example requires -hangroot-, which you can download
> by typing in Stata -ssc install hangroot-.
>
> Hope this helps,
> Maarten
>
> --------------------------
> Maarten L. Buis
> Institut fuer Soziologie
> Universitaet Tuebingen
> Wilhelmstrasse 36
> 72074 Tuebingen
> Germany
>
> http://www.maartenbuis.nl
> --------------------------
>
>
>
>
> *
> *   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/
>



--
--
Prakash Kashwan <http://mypage.iu.edu/~pkashwan/>
School of Public & Environmental Affairs (SPEA)
Workshop in Political Theory and Policy Analysis; Indiana University,
Bloomington
Research Fellow (2009) - International Foundation for Science (Sweden)

*
*   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/
Reply | Threaded
Open this post in threaded view
|  
Report Content as Inappropriate

Re: Suest v/s biprob in stata 11

Maarten buis
--- On Tue, 6/7/10, Prakash Kashwan wrote:
> I am conducting the two tests on same set of observations
> (so, the population is not divided into two groups).
> Moreover, in my case, I am using the same set of independent
> variables to estimate probabilities for two different
> dependent variables, and after I have run the -suest-
> command, I would like to figure out if the -suest-, did find
> significant correlation between the two error matrices, and
> if it makes sense for me to use -suest- in this case.

As I understand -suest- (see my previous post) it does not make
sense to use -suest- for such an hypothesis. To continue the
analogy, that would be like using a model with -robust- standard
errors to test for heteroskedasticity: the entire idea of the
model is that it does not estimate the heteroskedasticity and
that it does not need to. The advantage is that it is less
sensitive to wrong assumptions about heteroskedasticity, but the
price you must pay is that you can't use that model to check for
that either. Similarly, -suest- does not need these correlations
between error terms to get correct inference about the other
parameters, but the price is that you can't use -suest- to check
for such correlation.

-- Maarten

--------------------------
Maarten L. Buis
Institut fuer Soziologie
Universitaet Tuebingen
Wilhelmstrasse 36
72074 Tuebingen
Germany

http://www.maartenbuis.nl
--------------------------


     

*
*   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/
Loading...