SPSSX Discussion

estimating restricted systems of equations with SPSS

Classic

List

Threaded

9 messages Options

roxana dragan

estimating restricted systems of equations with SPSS

Hello,

I have a (panel/longitudinal) data set on a set of OECD countries over the
period of time 1970-2000. I have a lot of variables: V1, V2, V3.. up to V49.
The data looks like this:

country_1 1970 V1 V2 V3 V4 V5 V6....V49
country_1 1980 V1 V2 V3 V4 V5 V6....V49
country_1 1990 V1 V2 V3 V4 V5 V6....V49
country_1 2000 V1 V2 V3 V4 V5 V6....V49
country_2 1970 V1 V2 V3 V4 V5 V6....V49
country_2 1980 V1 V2 V3 V4 V5 V6....V49
country_2 1990 V1 V2 V3 V4 V5 V6....V49
country_2 2000 V1 V2 V3 V4 V5 V6....V49
country_3 1970 V1 V2 V3 V4 V5 V6....V49
country_3 1980......

The data are stacked. The data for a country are grouped together and
ordered by the time period. What I need to do is to estimate systems of
regressions like this:

V1 = constant1 + alpha*V3 + error1
V2 = constant2 + alpha*V3 + error2

The above two equations form a restricted system of equations. I want to
restrict "alpha" to be the same in both equations. I would need to estimate
restricted systems like this lots of times, not just a few times. For a
moment I thought I can "trick" the program by creating another data set
where I would stack V1 on top of V2 and V3 on top of itself and then just
run a simple regression - it would have been costly to that a lot of times
anyways, but I realized I would restrict the intercepts as well - I do not
want that - it was wrong. Currently I am banging my head against the GLM
command, trying to understand it. Do you have any advise on how to approach
my restricted estimation problem? Even an advise like keep on banging your
head against the GLM command would be helpful, at least I would know I am on
the right track.

Thanks,
Roxana

=====================
To manage your subscription to SPSSX-L, send a message to
[hidden email] (not to SPSSX-L), with no body text except the
command. To leave the list, send the command
SIGNOFF SPSSX-L
For a list of commands to manage subscriptions, send the command
INFO REFCARD

Peck, Jon

Re: estimating restricted systems of equations with SPSS

I didn't see an answer to this question posted. You can estimate your restricted system by stacking the data and using ordinary regression. In order to allow for separate constant terms, you would just calculate a separate country specific dummy variable and include these in the regression.

E.g., compute country_1 = country = "country+1" (or however the countries are identified).

Then include N-1 of these dummies in the equation along with the other variable(s).

However, you might be better off using the Mixed procedure to allow for the error structure likely to be present in panel data, and then you could specify the panel structure of the data directly in the procedure.

HTH,
Jon Peck

-----Original Message-----
From: SPSSX(r) Discussion [mailto:[hidden email]] On Behalf Of Roxana Dragan
Sent: Friday, May 16, 2008 2:01 PM
To: [hidden email]
Subject: [SPSSX-L] estimating restricted systems of equations with SPSS

Hello,

I have a (panel/longitudinal) data set on a set of OECD countries over the
period of time 1970-2000. I have a lot of variables: V1, V2, V3.. up to V49.
The data looks like this:

country_1 1970 V1 V2 V3 V4 V5 V6....V49
country_1 1980 V1 V2 V3 V4 V5 V6....V49
country_1 1990 V1 V2 V3 V4 V5 V6....V49
country_1 2000 V1 V2 V3 V4 V5 V6....V49
country_2 1970 V1 V2 V3 V4 V5 V6....V49
country_2 1980 V1 V2 V3 V4 V5 V6....V49
country_2 1990 V1 V2 V3 V4 V5 V6....V49
country_2 2000 V1 V2 V3 V4 V5 V6....V49
country_3 1970 V1 V2 V3 V4 V5 V6....V49
country_3 1980......

The data are stacked. The data for a country are grouped together and
ordered by the time period. What I need to do is to estimate systems of
regressions like this:

V1 = constant1 + alpha*V3 + error1
V2 = constant2 + alpha*V3 + error2

The above two equations form a restricted system of equations. I want to
restrict "alpha" to be the same in both equations. I would need to estimate
restricted systems like this lots of times, not just a few times. For a
moment I thought I can "trick" the program by creating another data set
where I would stack V1 on top of V2 and V3 on top of itself and then just
run a simple regression - it would have been costly to that a lot of times
anyways, but I realized I would restrict the intercepts as well - I do not
want that - it was wrong. Currently I am banging my head against the GLM
command, trying to understand it. Do you have any advise on how to approach
my restricted estimation problem? Even an advise like keep on banging your
head against the GLM command would be helpful, at least I would know I am on
the right track.

Thanks,
Roxana

=====================
To manage your subscription to SPSSX-L, send a message to
[hidden email] (not to SPSSX-L), with no body text except the
command. To leave the list, send the command
SIGNOFF SPSSX-L
For a list of commands to manage subscriptions, send the command
INFO REFCARD

=====================
To manage your subscription to SPSSX-L, send a message to
[hidden email] (not to SPSSX-L), with no body text except the
command. To leave the list, send the command
SIGNOFF SPSSX-L
For a list of commands to manage subscriptions, send the command
INFO REFCARD

roxana dragan

Re: estimating restricted systems of equations with SPSS

Jon,

Thank you for your reply. I was planning to use country dummies anyways, but
I do not see how that would solve my problem. The country dummy just
averages the residuals over a particular country for an equation. Now that
you brought that up, I think I can just average over the residuals of the
first equation and add that to the intercept of the stacked data equation to
get the intercept of the first equation. I think it is correct, but I am
just insecure and I need some confirmation. I would still need to estimate
standard errors for that new intercept - I have no clue how to do that. I
think the F statistic would change. I am a bit confused. I guess I will give
up spss and try another software.

Roxana

On Mon, May 19, 2008 at 8:59 AM, Peck, Jon <[hidden email]> wrote:

> I didn't see an answer to this question posted. You can estimate your
> restricted system by stacking the data and using ordinary regression. In
> order to allow for separate constant terms, you would just calculate a
> separate country specific dummy variable and include these in the
> regression.
>
> E.g., compute country_1 = country = "country+1" (or however the countries
> are identified).
>
> Then include N-1 of these dummies in the equation along with the other
> variable(s).
>
> However, you might be better off using the Mixed procedure to allow for the
> error structure likely to be present in panel data, and then you could
> specify the panel structure of the data directly in the procedure.
>
> HTH,
> Jon Peck
>
> -----Original Message-----
> From: SPSSX(r) Discussion [mailto:[hidden email]] On Behalf Of
> Roxana Dragan
> Sent: Friday, May 16, 2008 2:01 PM
> To: [hidden email]
> Subject: [SPSSX-L] estimating restricted systems of equations with SPSS
>
> Hello,
>
> I have a (panel/longitudinal) data set on a set of OECD countries over the
> period of time 1970-2000. I have a lot of variables: V1, V2, V3.. up to
> V49.
> The data looks like this:
>
> country_1 1970 V1 V2 V3 V4 V5 V6....V49
> country_1 1980 V1 V2 V3 V4 V5 V6....V49
> country_1 1990 V1 V2 V3 V4 V5 V6....V49
> country_1 2000 V1 V2 V3 V4 V5 V6....V49
> country_2 1970 V1 V2 V3 V4 V5 V6....V49
> country_2 1980 V1 V2 V3 V4 V5 V6....V49
> country_2 1990 V1 V2 V3 V4 V5 V6....V49
> country_2 2000 V1 V2 V3 V4 V5 V6....V49
> country_3 1970 V1 V2 V3 V4 V5 V6....V49
> country_3 1980......
>
> The data are stacked. The data for a country are grouped together and
> ordered by the time period. What I need to do is to estimate systems of
> regressions like this:
>
> V1 = constant1 + alpha*V3 + error1
> V2 = constant2 + alpha*V3 + error2
>
> The above two equations form a restricted system of equations. I want to
> restrict "alpha" to be the same in both equations. I would need to estimate
> restricted systems like this lots of times, not just a few times. For a
> moment I thought I can "trick" the program by creating another data set
> where I would stack V1 on top of V2 and V3 on top of itself and then just
> run a simple regression - it would have been costly to that a lot of times
> anyways, but I realized I would restrict the intercepts as well - I do not
> want that - it was wrong. Currently I am banging my head against the GLM
> command, trying to understand it. Do you have any advise on how to approach
> my restricted estimation problem? Even an advise like keep on banging your
> head against the GLM command would be helpful, at least I would know I am
> on
> the right track.
>
> Thanks,
> Roxana
>
> =====================
> To manage your subscription to SPSSX-L, send a message to
> [hidden email] (not to SPSSX-L), with no body text except the
> command. To leave the list, send the command
> SIGNOFF SPSSX-L
> For a list of commands to manage subscriptions, send the command
> INFO REFCARD
>

=====================
To manage your subscription to SPSSX-L, send a message to
[hidden email] (not to SPSSX-L), with no body text except the
command. To leave the list, send the command
SIGNOFF SPSSX-L
For a list of commands to manage subscriptions, send the command
INFO REFCARD

Peck, Jon

Re: estimating restricted systems of equations with SPSS

Country dummies provide country-specific intercepts. That’s what you wanted. But if you use country as a factor in GLM or, better, perhaps, in MIXED, where you can allow for the error structure, this will be taken care of automatically.

Depending on your data, you might at least want to allow the error variance to be country specific (diagonal covariance), which you can do with MIXED. With the large variation in the size of the OECD countries economies, constant variance across countries is probably unrealistic with typical economic dependent variables.

You can estimate those variances with MIXED and see how it works.

HTH,

Jon Peck

________________________________

From: Roxana Dragan [mailto:[hidden email]]
Sent: Tuesday, May 20, 2008 9:29 AM
To: Peck, Jon
Cc: [hidden email]
Subject: Re: [SPSSX-L] estimating restricted systems of equations with SPSS

Jon,

Thank you for your reply. I was planning to use country dummies anyways, but I do not see how that would solve my problem. The country dummy just averages the residuals over a particular country for an equation. Now that you brought that up, I think I can just average over the residuals of the first equation and add that to the intercept of the stacked data equation to get the intercept of the first equation. I think it is correct, but I am just insecure and I need some confirmation. I would still need to estimate standard errors for that new intercept - I have no clue how to do that. I think the F statistic would change. I am a bit confused. I guess I will give up spss and try another software.

Roxana

On Mon, May 19, 2008 at 8:59 AM, Peck, Jon <[hidden email]> wrote:

I didn't see an answer to this question posted. You can estimate your restricted system by stacking the data and using ordinary regression. In order to allow for separate constant terms, you would just calculate a separate country specific dummy variable and include these in the regression.

E.g., compute country_1 = country = "country+1" (or however the countries are identified).

Then include N-1 of these dummies in the equation along with the other variable(s).

However, you might be better off using the Mixed procedure to allow for the error structure likely to be present in panel data, and then you could specify the panel structure of the data directly in the procedure.

HTH,
Jon Peck

-----Original Message-----
From: SPSSX(r) Discussion [mailto:[hidden email]] On Behalf Of Roxana Dragan
Sent: Friday, May 16, 2008 2:01 PM
To: [hidden email]
Subject: [SPSSX-L] estimating restricted systems of equations with SPSS

Hello,

I have a (panel/longitudinal) data set on a set of OECD countries over the
period of time 1970-2000. I have a lot of variables: V1, V2, V3.. up to V49.
The data looks like this:

country_1 1970 V1 V2 V3 V4 V5 V6....V49
country_1 1980 V1 V2 V3 V4 V5 V6....V49
country_1 1990 V1 V2 V3 V4 V5 V6....V49
country_1 2000 V1 V2 V3 V4 V5 V6....V49
country_2 1970 V1 V2 V3 V4 V5 V6....V49
country_2 1980 V1 V2 V3 V4 V5 V6....V49
country_2 1990 V1 V2 V3 V4 V5 V6....V49
country_2 2000 V1 V2 V3 V4 V5 V6....V49
country_3 1970 V1 V2 V3 V4 V5 V6....V49
country_3 1980......

The data are stacked. The data for a country are grouped together and
ordered by the time period. What I need to do is to estimate systems of
regressions like this:

V1 = constant1 + alpha*V3 + error1
V2 = constant2 + alpha*V3 + error2

The above two equations form a restricted system of equations. I want to
restrict "alpha" to be the same in both equations. I would need to estimate
restricted systems like this lots of times, not just a few times. For a
moment I thought I can "trick" the program by creating another data set
where I would stack V1 on top of V2 and V3 on top of itself and then just
run a simple regression - it would have been costly to that a lot of times
anyways, but I realized I would restrict the intercepts as well - I do not
want that - it was wrong. Currently I am banging my head against the GLM
command, trying to understand it. Do you have any advise on how to approach
my restricted estimation problem? Even an advise like keep on banging your
head against the GLM command would be helpful, at least I would know I am on
the right track.

Thanks,
Roxana

=====================
To manage your subscription to SPSSX-L, send a message to
[hidden email] (not to SPSSX-L), with no body text except the
command. To leave the list, send the command
SIGNOFF SPSSX-L
For a list of commands to manage subscriptions, send the command
INFO REFCARD

roxana dragan

Re: estimating restricted systems of equations with SPSS

Jon,

stacking the data seems to be the solution. Than the problem I run into is
that I need to that a lot of times. For every single regression I need to
create a new data set. I have lots of regressions like that, 36 to be exact.
I am not looking forward to create 36 new data files.

Roxana

On Tue, May 20, 2008 at 10:56 AM, Peck, Jon <[hidden email]> wrote:

> Country dummies provide country-specific intercepts. That's what you
> wanted. But if you use country as a factor in GLM or, better, perhaps, in
> MIXED, where you can allow for the error structure, this will be taken care
> of automatically.
>
>
>
> Depending on your data, you might at least want to allow the error variance
> to be country specific (diagonal covariance), which you can do with MIXED.
> With the large variation in the size of the OECD countries economies,
> constant variance across countries is probably unrealistic with typical
> economic dependent variables.
>
>
>
> You can estimate those variances with MIXED and see how it works.
>
>
>
> HTH,
>
> Jon Peck
>
>
> ------------------------------
>
> *From:* Roxana Dragan [mailto:[hidden email]]
> *Sent:* Tuesday, May 20, 2008 9:29 AM
> *To:* Peck, Jon
> *Cc:* [hidden email]
> *Subject:* Re: [SPSSX-L] estimating restricted systems of equations with
> SPSS
>
>
>
> Jon,
>
> Thank you for your reply. I was planning to use country dummies anyways,
> but I do not see how that would solve my problem. The country dummy just
> averages the residuals over a particular country for an equation. Now that
> you brought that up, I think I can just average over the residuals of the
> first equation and add that to the intercept of the stacked data equation to
> get the intercept of the first equation. I think it is correct, but I am
> just insecure and I need some confirmation. I would still need to estimate
> standard errors for that new intercept - I have no clue how to do that. I
> think the F statistic would change. I am a bit confused. I guess I will give
> up spss and try another software.
>
> Roxana
>
> On Mon, May 19, 2008 at 8:59 AM, Peck, Jon <[hidden email]> wrote:
>
> I didn't see an answer to this question posted. You can estimate your
> restricted system by stacking the data and using ordinary regression. In
> order to allow for separate constant terms, you would just calculate a
> separate country specific dummy variable and include these in the
> regression.
>
> E.g., compute country_1 = country = "country+1" (or however the countries
> are identified).
>
> Then include N-1 of these dummies in the equation along with the other
> variable(s).
>
> However, you might be better off using the Mixed procedure to allow for the
> error structure likely to be present in panel data, and then you could
> specify the panel structure of the data directly in the procedure.
>
> HTH,
> Jon Peck
>
>
> -----Original Message-----
> From: SPSSX(r) Discussion [mailto:[hidden email]] On Behalf Of
> Roxana Dragan
> Sent: Friday, May 16, 2008 2:01 PM
> To: [hidden email]
> Subject: [SPSSX-L] estimating restricted systems of equations with SPSS
>
> Hello,
>
> I have a (panel/longitudinal) data set on a set of OECD countries over the
> period of time 1970-2000. I have a lot of variables: V1, V2, V3.. up to
> V49.
> The data looks like this:
>
> country_1 1970 V1 V2 V3 V4 V5 V6....V49
> country_1 1980 V1 V2 V3 V4 V5 V6....V49
> country_1 1990 V1 V2 V3 V4 V5 V6....V49
> country_1 2000 V1 V2 V3 V4 V5 V6....V49
> country_2 1970 V1 V2 V3 V4 V5 V6....V49
> country_2 1980 V1 V2 V3 V4 V5 V6....V49
> country_2 1990 V1 V2 V3 V4 V5 V6....V49
> country_2 2000 V1 V2 V3 V4 V5 V6....V49
> country_3 1970 V1 V2 V3 V4 V5 V6....V49
> country_3 1980......
>
> The data are stacked. The data for a country are grouped together and
> ordered by the time period. What I need to do is to estimate systems of
> regressions like this:
>
> V1 = constant1 + alpha*V3 + error1
> V2 = constant2 + alpha*V3 + error2
>
> The above two equations form a restricted system of equations. I want to
> restrict "alpha" to be the same in both equations. I would need to estimate
> restricted systems like this lots of times, not just a few times. For a
> moment I thought I can "trick" the program by creating another data set
> where I would stack V1 on top of V2 and V3 on top of itself and then just
> run a simple regression - it would have been costly to that a lot of times
> anyways, but I realized I would restrict the intercepts as well - I do not
> want that - it was wrong. Currently I am banging my head against the GLM
> command, trying to understand it. Do you have any advise on how to approach
> my restricted estimation problem? Even an advise like keep on banging your
> head against the GLM command would be helpful, at least I would know I am
> on
> the right track.
>
> Thanks,
> Roxana
>
> =====================
> To manage your subscription to SPSSX-L, send a message to
> [hidden email] (not to SPSSX-L), with no body text except the
> command. To leave the list, send the command
> SIGNOFF SPSSX-L
> For a list of commands to manage subscriptions, send the command
> INFO REFCARD
>
>
>

Peck, Jon

Re: estimating restricted systems of equations with SPSS

Why would you need a new dataset for each regression? You can merge them all together in one command. Just use ADD FILES to merge them and then work with the combined dataset. You can do up to 49 in a single command. If you want to work with an individual country or individual year, use the filter capabilities to select the cases you need based on your country or year variables.

________________________________

From: Roxana Dragan [mailto:[hidden email]]
Sent: Tuesday, May 20, 2008 10:24 AM
To: Peck, Jon
Cc: [hidden email]
Subject: Re: [SPSSX-L] estimating restricted systems of equations with SPSS

Jon,

stacking the data seems to be the solution. Than the problem I run into is that I need to that a lot of times. For every single regression I need to create a new data set. I have lots of regressions like that, 36 to be exact. I am not looking forward to create 36 new data files.

Roxana

On Tue, May 20, 2008 at 10:56 AM, Peck, Jon <[hidden email]> wrote:

Country dummies provide country-specific intercepts. That's what you wanted. But if you use country as a factor in GLM or, better, perhaps, in MIXED, where you can allow for the error structure, this will be taken care of automatically.

Depending on your data, you might at least want to allow the error variance to be country specific (diagonal covariance), which you can do with MIXED. With the large variation in the size of the OECD countries economies, constant variance across countries is probably unrealistic with typical economic dependent variables.

You can estimate those variances with MIXED and see how it works.

HTH,

Jon Peck

________________________________

From: Roxana Dragan [mailto:[hidden email]]
Sent: Tuesday, May 20, 2008 9:29 AM
To: Peck, Jon
Cc: [hidden email]
Subject: Re: [SPSSX-L] estimating restricted systems of equations with SPSS

Jon,

Thank you for your reply. I was planning to use country dummies anyways, but I do not see how that would solve my problem. The country dummy just averages the residuals over a particular country for an equation. Now that you brought that up, I think I can just average over the residuals of the first equation and add that to the intercept of the stacked data equation to get the intercept of the first equation. I think it is correct, but I am just insecure and I need some confirmation. I would still need to estimate standard errors for that new intercept - I have no clue how to do that. I think the F statistic would change. I am a bit confused. I guess I will give up spss and try another software.

Roxana

On Mon, May 19, 2008 at 8:59 AM, Peck, Jon <[hidden email]> wrote:

I didn't see an answer to this question posted. You can estimate your restricted system by stacking the data and using ordinary regression. In order to allow for separate constant terms, you would just calculate a separate country specific dummy variable and include these in the regression.

E.g., compute country_1 = country = "country+1" (or however the countries are identified).

Then include N-1 of these dummies in the equation along with the other variable(s).

However, you might be better off using the Mixed procedure to allow for the error structure likely to be present in panel data, and then you could specify the panel structure of the data directly in the procedure.

HTH,
Jon Peck

-----Original Message-----
From: SPSSX(r) Discussion [mailto:[hidden email]] On Behalf Of Roxana Dragan
Sent: Friday, May 16, 2008 2:01 PM
To: [hidden email]
Subject: [SPSSX-L] estimating restricted systems of equations with SPSS

Hello,

I have a (panel/longitudinal) data set on a set of OECD countries over the
period of time 1970-2000. I have a lot of variables: V1, V2, V3.. up to V49.
The data looks like this:

country_1 1970 V1 V2 V3 V4 V5 V6....V49
country_1 1980 V1 V2 V3 V4 V5 V6....V49
country_1 1990 V1 V2 V3 V4 V5 V6....V49
country_1 2000 V1 V2 V3 V4 V5 V6....V49
country_2 1970 V1 V2 V3 V4 V5 V6....V49
country_2 1980 V1 V2 V3 V4 V5 V6....V49
country_2 1990 V1 V2 V3 V4 V5 V6....V49
country_2 2000 V1 V2 V3 V4 V5 V6....V49
country_3 1970 V1 V2 V3 V4 V5 V6....V49
country_3 1980......

The data are stacked. The data for a country are grouped together and
ordered by the time period. What I need to do is to estimate systems of
regressions like this:

V1 = constant1 + alpha*V3 + error1
V2 = constant2 + alpha*V3 + error2

The above two equations form a restricted system of equations. I want to
restrict "alpha" to be the same in both equations. I would need to estimate
restricted systems like this lots of times, not just a few times. For a
moment I thought I can "trick" the program by creating another data set
where I would stack V1 on top of V2 and V3 on top of itself and then just
run a simple regression - it would have been costly to that a lot of times
anyways, but I realized I would restrict the intercepts as well - I do not
want that - it was wrong. Currently I am banging my head against the GLM
command, trying to understand it. Do you have any advise on how to approach
my restricted estimation problem? Even an advise like keep on banging your
head against the GLM command would be helpful, at least I would know I am on
the right track.

Thanks,
Roxana

=====================
To manage your subscription to SPSSX-L, send a message to
[hidden email] (not to SPSSX-L), with no body text except the
command. To leave the list, send the command
SIGNOFF SPSSX-L
For a list of commands to manage subscriptions, send the command
INFO REFCARD

roxana dragan

Re: estimating restricted systems of equations with SPSS

Thanks a lot Jon.

On Tue, May 20, 2008 at 11:30 AM, Peck, Jon <[hidden email]> wrote:

> Why would you need a new dataset for each regression? You can merge them
> all together in one command. Just use ADD FILES to merge them and then work
> with the combined dataset. You can do up to 49 in a single command. If you
> want to work with an individual country or individual year, use the filter
> capabilities to select the cases you need based on your country or year
> variables.
>
>
> ------------------------------
>
> *From:* Roxana Dragan [mailto:[hidden email]]
> *Sent:* Tuesday, May 20, 2008 10:24 AM
>
> *To:* Peck, Jon
> *Cc:* [hidden email]
> *Subject:* Re: [SPSSX-L] estimating restricted systems of equations with
> SPSS
>
>
>
> Jon,
>
> stacking the data seems to be the solution. Than the problem I run into is
> that I need to that a lot of times. For every single regression I need to
> create a new data set. I have lots of regressions like that, 36 to be exact.
> I am not looking forward to create 36 new data files.
>
> Roxana
>
> On Tue, May 20, 2008 at 10:56 AM, Peck, Jon <[hidden email]> wrote:
>
> Country dummies provide country-specific intercepts. That's what you
> wanted. But if you use country as a factor in GLM or, better, perhaps, in
> MIXED, where you can allow for the error structure, this will be taken care
> of automatically.
>
>
>
> Depending on your data, you might at least want to allow the error variance
> to be country specific (diagonal covariance), which you can do with MIXED.
> With the large variation in the size of the OECD countries economies,
> constant variance across countries is probably unrealistic with typical
> economic dependent variables.
>
>
>
> You can estimate those variances with MIXED and see how it works.
>
>
>
> HTH,
>
> Jon Peck
>
>
> ------------------------------
>
> *From:* Roxana Dragan [mailto:[hidden email]]
> *Sent:* Tuesday, May 20, 2008 9:29 AM
> *To:* Peck, Jon
> *Cc:* [hidden email]
> *Subject:* Re: [SPSSX-L] estimating restricted systems of equations with
> SPSS
>
>
>
> Jon,
>
> Thank you for your reply. I was planning to use country dummies anyways,
> but I do not see how that would solve my problem. The country dummy just
> averages the residuals over a particular country for an equation. Now that
> you brought that up, I think I can just average over the residuals of the
> first equation and add that to the intercept of the stacked data equation to
> get the intercept of the first equation. I think it is correct, but I am
> just insecure and I need some confirmation. I would still need to estimate
> standard errors for that new intercept - I have no clue how to do that. I
> think the F statistic would change. I am a bit confused. I guess I will give
> up spss and try another software.
>
> Roxana
>
> On Mon, May 19, 2008 at 8:59 AM, Peck, Jon <[hidden email]> wrote:
>
> I didn't see an answer to this question posted. You can estimate your
> restricted system by stacking the data and using ordinary regression. In
> order to allow for separate constant terms, you would just calculate a
> separate country specific dummy variable and include these in the
> regression.
>
> E.g., compute country_1 = country = "country+1" (or however the countries
> are identified).
>
> Then include N-1 of these dummies in the equation along with the other
> variable(s).
>
> However, you might be better off using the Mixed procedure to allow for the
> error structure likely to be present in panel data, and then you could
> specify the panel structure of the data directly in the procedure.
>
> HTH,
> Jon Peck
>
>
> -----Original Message-----
> From: SPSSX(r) Discussion [mailto:[hidden email]] On Behalf Of
> Roxana Dragan
> Sent: Friday, May 16, 2008 2:01 PM
> To: [hidden email]
> Subject: [SPSSX-L] estimating restricted systems of equations with SPSS
>
> Hello,
>
> I have a (panel/longitudinal) data set on a set of OECD countries over the
> period of time 1970-2000. I have a lot of variables: V1, V2, V3.. up to
> V49.
> The data looks like this:
>
> country_1 1970 V1 V2 V3 V4 V5 V6....V49
> country_1 1980 V1 V2 V3 V4 V5 V6....V49
> country_1 1990 V1 V2 V3 V4 V5 V6....V49
> country_1 2000 V1 V2 V3 V4 V5 V6....V49
> country_2 1970 V1 V2 V3 V4 V5 V6....V49
> country_2 1980 V1 V2 V3 V4 V5 V6....V49
> country_2 1990 V1 V2 V3 V4 V5 V6....V49
> country_2 2000 V1 V2 V3 V4 V5 V6....V49
> country_3 1970 V1 V2 V3 V4 V5 V6....V49
> country_3 1980......
>
> The data are stacked. The data for a country are grouped together and
> ordered by the time period. What I need to do is to estimate systems of
> regressions like this:
>
> V1 = constant1 + alpha*V3 + error1
> V2 = constant2 + alpha*V3 + error2
>
> The above two equations form a restricted system of equations. I want to
> restrict "alpha" to be the same in both equations. I would need to estimate
> restricted systems like this lots of times, not just a few times. For a
> moment I thought I can "trick" the program by creating another data set
> where I would stack V1 on top of V2 and V3 on top of itself and then just
> run a simple regression - it would have been costly to that a lot of times
> anyways, but I realized I would restrict the intercepts as well - I do not
> want that - it was wrong. Currently I am banging my head against the GLM
> command, trying to understand it. Do you have any advise on how to approach
> my restricted estimation problem? Even an advise like keep on banging your
> head against the GLM command would be helpful, at least I would know I am
> on
> the right track.
>
> Thanks,
> Roxana
>
> =====================
> To manage your subscription to SPSSX-L, send a message to
> [hidden email] (not to SPSSX-L), with no body text except the
> command. To leave the list, send the command
> SIGNOFF SPSSX-L
> For a list of commands to manage subscriptions, send the command
> INFO REFCARD
>
>
>
>
>

roxana dragan

Re: estimating restricted systems of equations with SPSS

In reply to this post by Peck, Jon

Jon,

I noticed that the MIXED command uses maximum likelihood to estimate the
covariance parameters. What I really need is an OLS (ordinary least squares
= minimize over the sum of squared residuals) estimation of the fixed
effects. Now, it is true that in certain conditions the two estimation
procedures provide the same parameter estimates. For me, an ordinary user,
it is confusing and worrying. Maximum Likelihood has nice properties when
you have large samples, but finite small sample properties are usually not
known (Greene, "Econometric Analysis", 1993). Is it true that the parameter
estimates of the fixed effects in the MIXED procedure also uses maximum
likelihood estimation?

Thank you,
Roxana

On Tue, May 20, 2008 at 11:30 AM, Peck, Jon <[hidden email]> wrote:

Peck, Jon

Re: estimating restricted systems of equations with SPSS

Perhaps you should look at the algorithms document for MIXED, which is accessible from the SPSS Help menu. Maximum likelihood methods are generally Good Things. When their small sample properties are not known, it is unlikely that possibly simpler methods such as OLS are going to be available with known small-sample properties either.

Perhaps your advisor can offer further advice.

Regards,

Jon Peck

________________________________

From: Roxana Dragan [mailto:[hidden email]]
Sent: Wednesday, May 21, 2008 10:28 AM
To: Peck, Jon
Cc: [hidden email]
Subject: Re: [SPSSX-L] estimating restricted systems of equations with SPSS

Jon,

I noticed that the MIXED command uses maximum likelihood to estimate the covariance parameters. What I really need is an OLS (ordinary least squares = minimize over the sum of squared residuals) estimation of the fixed effects. Now, it is true that in certain conditions the two estimation procedures provide the same parameter estimates. For me, an ordinary user, it is confusing and worrying. Maximum Likelihood has nice properties when you have large samples, but finite small sample properties are usually not known (Greene, "Econometric Analysis", 1993). Is it true that the parameter estimates of the fixed effects in the MIXED procedure also uses maximum likelihood estimation?

Thank you,
Roxana