inclusion of dummies in a SUREG esimation

[ANTONIO NETO]

Hi,

In order to estimate a system of equations using SURE I tryed the XTSUR command by it has not converged ( 15 hours processing with No results). My datase is a unbalanced panel ( quartely observations on 48 financial conglomerates)


Alternatively I used dummies for quaters, years and conglomerates and regressed using the command REG3 - sure like a pooled cross section.

The results seems to make sense.

Dos anyone have some comments on my procedures, or recommend any test to verify if this is correcT (the procedure)?

Best regards,

Antonio

[Andrew Musau]

Hi Antonio... I did promise to have a look at your post. When you think of SURE, think of equation by equation OLS structure, therefore no dummies!! I would approach this in the following ways:

1) Find out why your model is not converging. Maybe start with a few variables and then build up slowly.
2) If unable to get XTSUR to converge, then consider estimating the model with sureg (or reg3, sure). As I indicated, with SUREG, you want the model to be specified equation by equation form. Therefore, reshape from long to wide. I show this with the Grunfeld dataset.

  
. webuse grunfeld

. summarize

    Variable |        Obs        Mean    Std. Dev.       Min        Max
-------------+---------------------------------------------------------
     company |        200         5.5    2.879489          1         10
        year |        200      1944.5    5.780751       1935       1954
      invest |        200    145.9583    216.8753        .93     1486.7
      mvalue |        200    1081.681     1314.47      58.12     6241.7
      kstock |        200    276.0172    301.1039         .8     2226.3
-------------+---------------------------------------------------------
        time |        200        10.5    5.780751          1         20

*IN THIS DATASET, YOU HAVE 10 COMPANIES OBSERVED OVER THE YEARS 1935-1954 (20 YEARS)

. reshape wide invest mvalue kstock, i(year) j(company)
(note: j = 1 2 3 4 5 6 7 8 9 10)

Data                               long   ->   wide
-----------------------------------------------------------------------------
Number of obs.                      200   ->      20
Number of variables                   6   ->      32
j variable (10 values)          company   ->   (dropped)
xij variables:
                                 invest   ->   invest1 invest2 ... invest10
                                 mvalue   ->   mvalue1 mvalue2 ... mvalue10
                                 kstock   ->   kstock1 kstock2 ... kstock10
-----------------------------------------------------------------------------

*AFTER RESHAPING, YOU HAVE 20 OBSERVATIONS, ONE FOR EACH YEAR.

You can now run SUREG, ideally by defining a loop if the number of variables is too large.

tsset year
forvalues i = 1/10 {
local vars  "`vars'  (mvalue`i' invest`i' kstock`i')"
}
sureg `vars'

Note there are no individual fixed effects because the intercept is now individual specific. Also, you can impose cross-equation restrictions if you wish. All this assumes that your panel is balanced. If it is unbalanced, Allen McDowell describes a method for estimating SURE with unbalanced equations using XTGEE. He does this step by step, so it should be easy to follow.

http://www.stata-journal.com/sjpdf.h...iclenum=st0079

[ANTONIO NETO]

Andrew,

Many thanks for your comments.

Sorry to go back to a basic question but I am new user of Stata.

When you say : "therefore no dummies" , does it mean that my procedure is incorrect?

best regards.

[Andrew Musau]

Yes. SURE is a multiple time series estimator and therefore your data should be in wide form if you want to use SUREG or REG3, sure. XTSUR was designed to handle panel data (in long form).

[ANTONIO NETO]

Ok. It seems to me that your sugestion of turnning my panel into a wide form is more comprehensible for my level. It is a pity that it is for balanced panel. I will try to adapt the alternative for unbalanced panel. Do you have other sugestions of examples?

many thanks

Antonio

[ANTONIO NETO]

Andrew,

An additional information: my panel is short.

[Andrew Musau]

Any additional example of using sureg with panel data (in long form) would replicate the procedure in #2. From #1, I was under the impression that you had a system of linear equations. However, looking at your earlier post (http://www.statalist.org/forums/foru...-demand-system), I now see that you have a system of non linear equations and therefore the linear estimator is not appropriate. Therefore, your question is how to estimate an Almost Ideal Demand System (AIDS) model with non linear SURE (nlsur). This is beyond my area of specialization, but a procedure is described at the bottom of page 14 onwards of the Stata manual.

http://www.stata.com/manuals13/rnlsur.pdf

You should download the dataset and study how the equations are specified, then adapt it to your case. Hopefully, someone who regularly deals with these types of models will read this post and come up with some additional suggestions.