Dear Kit,
I posted the original question. I appreciated all of those responses that
question generated and benefited from the ensuing discussions (even though I
may or may not agree with all of them). Your example program gave me some
idea how I could proceed with what I intended to do with Stata. (As a new
user, I've got a lot to learn from you guys.) Therefore, the purpose of my
question is well served. Thank you.
Nik's question as well as my original one, however, is legitimate. You may
think that "sureg" is a panel data estimator, but it does not handle fixed
effects by itself. Your example clearly indicated that you had to combine
"sureg" with "xtreg" to deal with the issue at hand, which was exactly what
I didn't but wanted to know about when I raised my question. Therefore, your
claim that by using "sureg," one is already estimating fixed effects is, I
hate to say, less than true.
Thank you again for your generous discussions and example program, even
though I wish that you had shared this example program with me right after I
posted my question.
Jeremy Z.
-----Original Message-----
From: [email protected]
[mailto:[email protected]]On Behalf Of baum
Sent: Wednesday, October 23, 2002 12:30 AM
To: Nick Winter; StataList
Subject: st: RE: Re: fixed effects and SUR
Nick,
If you take a panel and reshape wide by i, you can estimate the same model
on sureg, right? In my mind sureg is a panel data estimator.
. use http://fmwww.bc.edu/ec-p/data/Greene2000/TBL15-1
. tsset firm year
panel variable: firm, 1 to 5
time variable: year, 1935 to 1954
. xtreg i f c,fe
Fixed-effects (within) regression Number of obs =
100
Group variable (i) : firm Number of groups =
5
R-sq: within = 0.8003 Obs per group: min =
20
between = 0.7699 avg =
20.0
overall = 0.7782 max =
20
F(2,93) =
186.40
corr(u_i, Xb) = -0.1359 Prob > F =
0.0000
---------------------------------------------------------------------------
---
i | Coef. Std. Err. t P>|t| [95% Conf.
Interval]
-------------+-------------------------------------------------------------
---
f | .1059799 .015891 6.67 0.000 .0744236
.1375363
c | .3466596 .0241612 14.35 0.000 .2986803
.3946388
_cons | -62.59438 29.44191 -2.13 0.036 -121.0602
-4.128578
-------------+-------------------------------------------------------------
---
sigma_u | 120.02194
sigma_e | 69.117977
rho | .75095637 (fraction of variance due to u_i)
---------------------------------------------------------------------------
---
F test that all u_i=0: F(4, 93) = 58.96 Prob > F =
0.0000
. reshape wide i f c,i(year) j(firm)
(note: j = 1 2 3 4 5)
Data long -> wide
---------------------------------------------------------------------------
--
Number of obs. 100 -> 20
Number of variables 5 -> 16
j variable (5 values) firm -> (dropped)
xij variables:
i -> i1 i2 ... i5
f -> f1 f2 ... f5
c -> c1 c2 ... c5
---------------------------------------------------------------------------
--
. forv i=1/5 {
2. local rhs "`rhs' ( i`i' f`i' c`i') "
3. }
. di "`rhs'"
( i1 f1 c1) ( i2 f2 c2) ( i3 f3 c3) ( i4 f4 c4) ( i5 f5 c5)
. sureg `rhs'
Seemingly unrelated regression
----------------------------------------------------------------------
Equation Obs Parms RMSE "R-sq" chi2 P
----------------------------------------------------------------------
i1 20 2 84.94729 0.9207 261.3219 0.0000
i2 20 2 12.36322 0.9119 207.2128 0.0000
i3 20 2 26.46612 0.6876 46.88498 0.0000
i4 20 2 9.742303 0.7264 59.14585 0.0000
i5 20 2 95.85484 0.4220 14.9687 0.0006
----------------------------------------------------------------------
---------------------------------------------------------------------------
---
| Coef. Std. Err. z P>|z| [95% Conf.
Interval]
-------------+-------------------------------------------------------------
---
i1 |
f1 | .120493 .0216291 5.57 0.000 .0781007
.1628853
c1 | .3827462 .032768 11.68 0.000 .318522
.4469703
_cons | -162.3641 89.45922 -1.81 0.070 -337.7009
12.97279
-------------+-------------------------------------------------------------
---
i2 |
f2 | .0695456 .0168975 4.12 0.000 .0364271
.1026641
c2 | .3085445 .0258635 11.93 0.000 .2578529
.3592362
_cons | .5043112 11.51283 0.04 0.965 -22.06042
23.06904
-------------+-------------------------------------------------------------
---
i3 |
f3 | .0372914 .0122631 3.04 0.002 .0132561
.0613268
c3 | .130783 .0220497 5.93 0.000 .0875663
.1739997
_cons | -22.43892 25.51859 -0.88 0.379 -72.45443
27.57659
-------------+-------------------------------------------------------------
---
i4 |
f4 | .0570091 .0113623 5.02 0.000 .0347395
.0792788
c4 | .0415065 .0412016 1.01 0.314 -.0392472
.1222602
_cons | 1.088878 6.258805 0.17 0.862 -11.17815
13.35591
-------------+-------------------------------------------------------------
---
i5 |
f5 | .1014782 .0547837 1.85 0.064 -.0058958
.2088523
c5 | .3999914 .1277946 3.13 0.002 .1495186
.6504642
_cons | 85.42324 111.8774 0.76 0.445 -133.8525
304.6989
---------------------------------------------------------------------------
---
If you impose constraints that the slopes are common across equations of
the SUR, you're going to be very close to a xtreg,fe model.
Kit
--On Tuesday, October 22, 2002 11:10 -0400 Nick Winter
<[email protected]> wrote:
> Kit,
>
> I'm confused by your mentioning of -sureg- in relation to panel models.
> How does sureg estimate what you describe? I thought of -sureg- as
> estimating different models on the same cases; not the same models on
> different sets of cases?
>
> But maybe I'm just missing something obvious?
>
> Thanks
> Nick
>
>
> -----------------------------------------------------------
> Nicholas Winter, Ph.D. P 202.939.5343
> Policy Studies Associates F 202.939.5732
> 1718 Connecticut Avenue, NW [email protected]
> Washington, DC 20009-1148 www.policystudies.com
> -----------------------------------------------------------
>
>> -----Original Message-----
>> From: Kit Baum, Faculty Micro Resource Center [mailto:[email protected]]
>> Sent: Tuesday, October 22, 2002 10:57 AM
>> To: [email protected]
>> Subject: st: Re: fixed effects and SUR
>>
>>
>>
>> --On Tuesday, October 22, 2002 2:33 -0400 Jeremy wrote:
>>
>> >
>> > Hi Everybody,
>> >
>> > Could anybody give me some idea how to estimate SUR with
>> fixed effects
>> > using Stata? I'm new to Stata. All I know at this point is
>> that I could
>> > use XTREG to estimate a single equation with fixed effects,
>> and SUREG to
>> > estimate a system of equations. I've no idea how to proceed
>> from here.
>> > For your information, I'm trying to use Stata to "cross-check" some
>> > estimation results I obtained using TSP. Your answers will
>> be gratefully
>> > appreciated.
>> >
>> > Jeremy Z.
>> >
>> and cb23 responded
>> >
>> > Just in case no-one comes up with a correct answer on this,
>> I would try
>> > the following :-
>> >
>> > I think I am right in saying that the xtreg fixed effects
>> model is just
>> > a standard OLS model with dummies for the groups with one
>> alteration: in
>> > OLS we choose a baseline dummy for which we set the
>> coefficient to zero
>> > and in fixed effects we sum the dummy coefficients on all groups to
>> > zero. This means that the coefficients on the other Xs
>> should be the
>> > same in fixed effects and the dummy variable approach, and the only
>> > difference will be in the coefficients of the constant and the group
>> > dummies/effects. You should try this to make sure it
>> works. You can
>> > even try to choose the baseline dummy such that the
>> coefficients on the
>> > other dummies are close to summing to zero.
>> >
>> > Than, having realised we can roughly write a fixed effect model as a
>> > standard equation, you could then rewrite your fixed effect
>> equations in
>> > a dummy variable form and stick them in to a SUR model.
>>
>>
>> I find this quite confused. Note that if we start with the
>> most general
>> (infeasible) model of panel data, in which every i and t has its own
>> coefficient vector, we can define special cases:
>>
>> a) all slopes constant over i and t, s^2 constant over i and
>> t, intercept
>> varies over i
>>
>> b) intercept, slopes, and s^2 all have an i subscript, but
>> are constant
>> over t
>>
>> The former case is one-way (individual) fixed effects, aka
>> LSDV (dummy var)
>> model, which may be estimated by xtreg, fe or areg. Note that
>> normalisation
>> of the intercepts makes no difference here; no matter whether
>> you include a
>> constant and (n-1) dummies, or express data as demeaned by
>> individual, you
>> will get the same estimates in terms of significance.
>>
>> The latter case is Zellner SUR, estimable via sureg. This is a 'fixed
>> effect' model, in that each individual has his/her own
>> equation (thus N < T
>> for standard SUR), with his/her own intercept, set of slopes,
>> and s^2. One
>> can consider special cases of SUR in which further
>> constraints are imposed
>> (e.g. common slopes over units) which, since SUR is a GLS
>> estimator, takes
>> you back very close to individual fixed effects (except that
>> SUR allows for
>> s^2_i, whereas
>> IFE imposes a single s^2 on the entire panel).
>>
>> So I don't know what it means to estimate SUR with fixed
>> effects; if you're
>> using SUR, you are already estimating individual fixed
>> effects, and more.
>>
>> Kit
>> --------------------------------------------------------------------
>> Kit Baum, Faculty Micro Resource Center [email protected]
>> Academic Technology Services, Boston College http://www.bc.edu/ats
>> http://fmwww.bc.edu/FMRC/ http://fmwww.bc.edu/GStat/
>> *
>> * For searches and help try:
>> * http://www.stata.com/support/faqs/res/findit.html
>> * http://www.stata.com/support/statalist/faq
>> * http://www.ats.ucla.edu/stat/stata/
>>
>
*
* For searches and help try:
* http://www.stata.com/support/faqs/res/findit.html
* http://www.stata.com/support/statalist/faq
* http://www.ats.ucla.edu/stat/stata/
*
* For searches and help try:
* http://www.stata.com/support/faqs/res/findit.html
* http://www.stata.com/support/statalist/faq
* http://www.ats.ucla.edu/stat/stata/
