Re: st: fractional response using xtgee - correct implementation of random effects and fixed effectst?

[Nick Cox <njcoxstata@gmail.com>]

Thanks for this. I was not inferring or implying any intent to offend
at all, just making factual points -- more for others to read than
yourself.

But I think it would have been much better to mention that you did
know about Leslie Papke's code on the page I cited.

To the best of my knowledge Leslie Papke is not a member of Statalist
so if you have any specific questions about her do-files you may have
to ask her directly. I absolutely cannot speak for her, but my wild
guess would be that the do-files are there essentially to document
what was done in a particular paper. I can't advise what other
materials may be available.

I haven't applied your kind of analysis to panel data but others may
be able to add extra comments.

Nick
njcoxstata@gmail.com


On 4 July 2013 13:29,  <Asmund.Rygh@bi.no> wrote:
> Nick,
>
> thank you for your reply. First of all: I am sorry if my comment about
> "not answered yet" came across as some sort of criticism. My only reason
> for including it was
> to indicate that the specific information I need is not already out there
> on Statalist, in which case I would be wasting people's time to ask it
> again. Unfortunately, as you point out I also inserted
> the wrong link (the correct one is
> http://www.stata.com/statalist/archive/2012-01/msg00595.html).and
> therefore I realize it appeared that I missed the responses from Kelvin. I
> am certainly not
> implying that Kelvin's responses were not helpful: Indeed they were.
> However, they mainly concerned the issue of unbalanced panels (which is
> not an issue for me), rather than the implementation
> of random or fixed effects (at least as far as I can see). Anyway, I will
> try to be more precise next time I post something.
>
> Thank you also for the link to Papke's home page. I did actually already
> have a look at this webpage and the files before I posted my first mail,
> However,  just by looking at the code I am still a bit unsure about
> whether they
> include fixed or random effects. The relevant excerpt from the code (in
> the file math4_boot_exog_gee.do) posted on the webpage is, I guess:
>
> glm math4 lavgrexp alavgrexp lunch alunch lenroll alenroll y96-y01 if
> year>1994, fa(bin) link(probit) cluster(distid)
> mat b = e(b)
> xtgee math4 lavgrexp lunch lenroll alavgrexp alunch alenroll y96-y01,
> fa(bi) link(probit) corr(exch) robust from(b,skip)
>
> My guess would be that the first step (-glm-) of this code is the first
> step in the two-step procedure discussed in section 4 in the 2008 article.
> I would also guess that the a-prefixes on the variables here could refer
> to
> "average". This is why I wanted to get some more information about the
> files (but, as mentioned, I am not able to access
> https://www.msu.edu/~ec/faculty/papke/papke.html where I assume such
> information can be found). I am still also unsure
> about whether omitting variables such as "alunch" means that it will be a
> random effects model. Therefore, further advice would be much appreciated.


From:   Nick Cox <njcoxstata@gmail.com>

> The 2012 link you give here is to a discussion in 2013.
>
> Otherwise I went to Leslie Papke's home page
> http://econ.msu.edu/faculty/papke/ and found a link to Stata code
> there.
>
> As a kind of meta-comment it is important to realise that absolutely
> no mechanism ensures that Statalist questions will be answered. It is
> entirely up to individuals to decide what they wish to answer.
>
> In the particular case of
> http://www.stata.com/statalist/archive/2013-06/msg00434.html there
> were several helpful replies from Kelvin Mulungu.
>
> But "not answered yet", even when correct,  means no more than it
> says. It doesn't mean waiting in a queue.
>
> Nick
> njcoxstata@gmail.com
>
> On 4 July 2013 12:23,  <Asmund.Rygh@bi.no> wrote:
>> (First line)
>>
>> Dear Statalist members,
>>
>> I would like to model the degree of internationalization of firms (a
>> fractional variable bounded between 0 and 1) using a strongly balanced
>> panel of 30 firms over an 11-year period, and test both the use of fixed
>> and of random effects. I am using Stata 12.
>>
>> I have read Papke and Wooldridge's (2008) article Panel data methods for
>> fractional response variables... (see
>> http://www.sciencedirect.com/science/article/pii/S030440760800050X) and
>> consulted the archives of Statalist. I found one relevant thread from
> 2012
>> at http://www.stata.com/statalist/archive/2013-06/msg00517.html as well
> as
>> a very recent one (last post 12 June) at
>> http://www.stata.com/statalist/archive/2013-06/msg00434.html which does
>> not appear to have been answered yet.
>>
>> In the first link from 2012 it is suggested to go to the website of
> Leslie
>> Papke at https://www.msu.edu/~ec/faculty/papke/papke.html where one
> "would
>> find lots of invaluable information with detailed codes as to how
>> implement in case of panel data". Perhaps, the answers to my question
>> could be found here. However, I am not able to access this site and the
>> following message is returned: You don't have permission to access
>> /~ec/faculty/papke/papke.html on this server. Thus, my first question is
>> simply whether somebody knows if I can access the material in Papke's
>> webpage from somewhere else? I have made extensive searches on the web
> but
>> couldn't find anything.
>>
>> To my main question: From the discussion in the above threads (as well
> as
>> my experience with using random effects and fixed effects in Stata for
>> normal variables), I am sort of guessing that the following two commands
>> using -xtgee- will achieve what I want:
>>
>> Random effects: xtgee dependent_variable independent_variables,
>> family(binomial) link(logit) vce(robust)
>>
>> Fixed effects: xtgee dependent_variable independent_variables
>> time_averages_independent_variables, family(binomial) link(logit) robust
>>
>> In the first case, I set firm and year as the panel variables (-xtset-)
>> and the output states that the group variable is firm. In the second
> case,
>> which I did not test yet, my interpretation of the discussion in PW2008
>> (see p. 123) and at the above mentioned links is that including time
>> averages will give me the fixed effects. Is this correct?
*
*   For searches and help try:
*   http://www.stata.com/help.cgi?search
*   http://www.stata.com/support/faqs/resources/statalist-faq/
*   http://www.ats.ucla.edu/stat/stata/