Bookmark and Share

Notice: On March 31, it was announced that Statalist is moving from an email list to a forum. The old list will shut down at the end of May, and its replacement, statalist.org is already up and running.


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

Re: st: xtlogit: panel data transformation's recast to double makes model incomputable


From   Nick Cox <njcoxstata@gmail.com>
To   "statalist@hsphsun2.harvard.edu" <statalist@hsphsun2.harvard.edu>
Subject   Re: st: xtlogit: panel data transformation's recast to double makes model incomputable
Date   Wed, 3 Apr 2013 13:02:38 +0100

The link is no more than this: fitting the wrong model seems unlikely
to help you. Also, Statalist has been explained as a discussion list,
not a help line, from its outset; people are not obliged to focus
entirely on the question being asked, but feel free to comment on
other aspects.

My impression from a distance is that the very skewed distributions of
your predictors pose possibly your biggest problem, as I said on Cross
Validated.

On 3 April 2013 12:54, Tom <tommedema@gmail.com> wrote:
> I understand what you mean. Note that I did not use any of these other
> models until unexpected computational problems appeared.
>
> I chose for a fixed effects model because I do not expect prices to be
> determined randomly, but rather to be inherent to the underlying
> company (each group represents a company). I expect each company to
> have a (time independent) effect on its prices. Therefore a fixed
> effects model seems appropriate. Also, the initial estimates are often
> significant and even the simpler model seem to have explanatory and
> predictive power.
>
> You mention that I can decide on another correlation structure for the
> observations, but according to the -xtlogit- help this is only
> applicable for population average (PA) models.
>
> I still don't quite see how this could cause the behavior that I'm
> seeing though. There are even issues with some independent variables
> when they are the only explanatory variable in the model, as shown in
> my earlier posts (e.g. -xtlogit depc_gpf30 close, fe-).
>
> I'm currently trying to determine if any of the companies have extreme
> leverage effects that could cause some of the issues.
>
> Tom
>
> On Wed, Apr 3, 2013 at 1:04 PM, Nick Cox <njcoxstata@gmail.com> wrote:
>> My intentions are to help you think about what you are doing.
>>
>> I don't think I can advise at all on what's a realistic data
>> generation process for your data and therefore I don't have extra
>> suggestions for your analyses. Quite apart from anything else, I don't
>> think you have said anything much about the data and what underlies
>> them, but I've seen enough to know that I am not working on similar
>> data.
>>
>> It is difficult to keep track of the various models that you have
>> tried, including -logit-, -xtlogit-, -clogit- and perhaps others.
>> Ideally, you should not be trying  for whatever will fit, but choosing
>> a model because the associated generation process matches your data
>> and the problem.
>>
>> I am asking what are you assuming about the dependence structure of
>> your data within panels, and your answer implies to me that you
>> haven't thought about it and/or are not aware of the associated
>> statistical issues. (Econometric issues, if you prefer.) By choosing
>> -xtlogit, fe- you make one set of choices that way, but as the help
>> for -xtlogit- explains, other choices allow differing assumptions
>> about correlation structure.
>>
>> Nick
>>
>> On 3 April 2013 11:38, Tom <tommedema@gmail.com> wrote:
>>> "Your various models all
>>> appear to be implying that there is no dependence other than implied
>>> by the panel structure, that is to say that prices at successive dates
>>> for the same panel are mutually independent, which seems a very strong
>>> assumption to me."
>>>
>>> This is not my intention. I've set the within-group date for this
>>> exact reason. Why do you suspect that I am assuming such independence
>>> among dates within the same group?
>>>
>>> Or mayhaps your intentions will become clear to me if you tell me what
>>> you'd suggest me to do. At the moment I'm not quite sure.
>>>
>>> Thank you. I will use the - character from now on to indicate commands.
>>>
>>> Tom
>>>
>>> On Wed, Apr 3, 2013 at 11:46 AM, Nick Cox <njcoxstata@gmail.com> wrote:
>>>> I think you are missing my point. Time series have some kind of
>>>> dependence structure, unless they happen to be white noise. Ignoring
>>>> that dependence won't make it disappear. Your various models all
>>>> appear to be implying that there is no dependence other than implied
>>>> by the panel structure, that is to say that prices at successive dates
>>>> for the same panel are mutually independent, which seems a very strong
>>>> assumption to me.
>>>>
>>>> A tiny point of presentation, yet another covered in the FAQ. Left
>>>> ticks around command names such as `tsset` have no special effects on
>>>> Statalist; the convention recommended is exemplified by -tsset-.
>>>>
>>>> Nick
>>>>
>>>> On 3 April 2013 10:13, Tom <tommedema@gmail.com> wrote:
>>>>> Hi Nick,
>>>>>
>>>>> They are time series. `tsset` returns:
>>>>>
>>>>> panel variable:  ticker_id (unbalanced)
>>>>> time variable:  date, 101 to 532
>>>>> delta:  1 unit
>>>>>
>>>>> Each group has its own continuous time frame, ranging from 101 to 532
>>>>> (the first 100 dates have been dropped because they were required to
>>>>> create lagged variables).
>>>>>
>>>>> The only issue that I could possible see arising from time series is
>>>>> that the dates are not "synchronised" inbetween groups. With this I
>>>>> mean that t == 200 is a different point in time for id == 1 than t ==
>>>>> 200 for the group with id == 2.
>>>>>
>>>>> But I don't think time series are the issue because clogit is not
>>>>> using them (it only uses the panel id variable), and it is returning
>>>>> exactly the same results as xtlogit.
>>>>>
>>>>> I was hoping that something obvious would arise from the many log
>>>>> files I submitted a couple of posts back :)
>>>>>
>>>>> Tom
>>>>>
>>>>>
>>>>> On Wed, Apr 3, 2013 at 10:04 AM, Nick Cox <njcoxstata@gmail.com> wrote:
>>>>>> As if there weren't enough problems, what is the assumption here about
>>>>>> time series structure? I may have missed something in a complicated
>>>>>> thread, but the original data look like time series to me.
>> *
>> *   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/
> *
> *   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/



-- 
Nick
njcoxstata@gmail.com
*
*   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/


© Copyright 1996–2014 StataCorp LP   |   Terms of use   |   Privacy   |   Contact us   |   Site index