Statalist


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

RE: st: Stata 9.2 versus Limdep


From   "Nick Cox" <n.j.cox@durham.ac.uk>
To   <statalist@hsphsun2.harvard.edu>
Subject   RE: st: Stata 9.2 versus Limdep
Date   Thu, 26 Mar 2009 18:23:03 -0000

Unfortunately, I think that the spirit of Michael's comments still
applies. 

A vague but not very helpful answer is that evidently Stata and Limdep
are not using exactly the same data, so some difference in results is
not at all surprising. 

Of your three points, #1 clearly has no bearing on any difference
between Stata and Limdep here. I don't understand all of #2 but the same
appears to be true. 

#3 raises the further question: How would Limdep not ignore missing
values? What would it do if the SKIP were not specified? But the
important question is still whether using SKIP has the same implications
as Stata's default treatment of missing values. 

This can really only be explored properly if other people with access to
both Stata 9.2 and Limdep and their documentation can experiment with 

1. Exactly the same dataset 

And 

2. Exactly the same commands as you typed, without any omissions
whatsoever. 

Otherwise, I don't see that experts in this field (not me) have enough
hard to bite on. 

As often mentioned on this list, one answer to #1 is to use a standard
dataset downloadable from within Stata, run that and export it to
Limdep; or vice versa, if such a thing is possible. 

It can be much easier to explore these questions via Stata technical
support, but I can't answer for whether StataCorp has a copy of the
version of Limdep you have (or indeed any other). 

You are also presuming that people competent to answer this question
recognise all the references you give in name (date) form. That is
likely to be true, but Statalist protocol is to expect full references. 

Nick 
n.j.cox@durham.ac.uk 

Marc Goergen

I would like to provide you with further information about my query (see
below for Michael's comments/queries).

1. I have always used Limdep and have only recently switched to Stata as
Stata seems to be the only package able to estimate the Blundell-Bond
(1998) system GMM estimator for a dynamic panel data model.

2. Bond (2002) explains the typical approach in determining whether the
first-differenced GMM estimator (Arellano and Bond (1991) or the system
GMM estimator is preferable for a particular model and dataset. An AR(1)
regression is run using a pooled OLS regression and a Within Groups
regression. Without going into too much detail, due to the omitted
variable bias, the coefficient on the lagged dependent variable obtained
from the pooled OLS regression tends to be upward biased and serves as
an upper bound. Conversely, Within Groups tends to produce downward
biased estimators. This is the reason why I used REGRESS and estimated a
simple, pooled OLS regression.

3. SKIP is a command used in Limdep to ignore missing values. This is
not done automatically and can create avoc.


>>> "Michael I. Lichter" <MLichter@Buffalo.EDU> 26/03/2009 16:03 >>>

That's not enough information for you to get an answer. What's special 
about your dependent variable that led you to use LIMDEP? You're using 
regular OLS regression in Stata; were you doing something different 
there? What does SKIP mean? You can't assume that anybody here knows 
enough about LIMDEP to interpret this. If you have fewer cases in Stata,

it's because something is missing. Were you doing pairwise deletion in 
LIMDEP? Also, if this is panel data, why are you using regress instead 
of xtreg? Explain this on the list, not to me directly.

Marc Goergen wrote:

> I have been running a pooled OLS regression on the same dataset in
both Limdep and Stata 9.2. However, the results I obtained are
significantly different:
>
> 1. Limdep seems to include roughly 10% more observations than Stata
does, despite using the skip command in Limdep.
>
> 2. Some of the variables that were highly significant in Limdep are no
longer significant in Stata. This is especially true for a dummy
variable which was significant at the 0.1% level in Limdep, but at best
has a p-value of 0.39 in Stata.
>
> 3. While for the case of Limdep the results for OLS without group
dummy variables are significantly different from those with group dummy
variables, in Stata the differences are much less pronounced.
>
> The command I used in Limdep is:
>
> REGRESS; LHS= LNAF; RHS= LNAF[-1],LNTA[-1], ...; STR=CODE;
PERIOD=YEAR; PANEL$
>
> The command I used in Stata is:
>
> regress lnaf l.lnaf l.lnta ... , robust cluster(code)
>
> Dropping the "robust" and "cluster" options in Stata does not bring
the results more in line with those obtained from Limdep.


*
*   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   |   What's new   |   Site index