Statalist


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

Re: st: clogit, "initial values not feasible" error


From   sjsamuels@gmail.com
To   statalist@hsphsun2.harvard.edu
Subject   Re: st: clogit, "initial values not feasible" error
Date   Fri, 18 Dec 2009 10:04:35 -0500

Okay, I think that I understand what you are trying to do.   Each
combination of predictors--x1 x2 x3 province, sector, year, country
defines a denominator "population" (which could have n = 1) if your
outcome is "binary" instead of "binomial").   For each combination
your outcome event is: there were positive exports (from the province
to the country) or not.

Ff all predictors are categorical, for example, you w have a multi-way
table with a zero or one assigned to each cell, assuming a binary
outcome.  Some cells might be completely empty if there are no
observations for the particular predictor combination.  If so, this
would be notationally a legitimate model and suppressing the country
coefficients with -clogit- would be okay.

The difficulties I see in the model are that it assumes 1) there are
no interactions;  and 2) that observations in different cells are
independent. These assumptions would be violated, for example, if
there were certain favored trading combinations of provinces and
countries.

Steve


On Fri, Dec 18, 2009 at 6:07 AM, joachim jarreau
<joachim.jarreau@gmail.com> wrote:
> Thank you for your answer. I have good reasons to think that
> heterogeneity across countries also matters for the outcome i study,
> which is why Id also want to control for it using country
> fixed-effects. I only defined "country" as the "group" in -clogit- in
> order to minimize the number of fixed effects I include "by hand" -
> because there is more countries than provinces or sectors. I thought
> that this would reduce computation time, and that it was equivalent to
> defining e.g. provinces as the group variable. It seems to me that it
> should lead to the same model being estimated.. but I may be wrong in
> making this assumption - Ill check this.
>
> joachim
>
>
> On Thu, Dec 17, 2009 at 10:10 PM,  <sjsamuels@gmail.com> wrote:
>> I don't know if this is related to your issue. Probabilistically,
>> events originate in a  province, not in the country exported to.  If
>> you have multiple observations per province, it appears to me that
>> "province" should be the group variable for -clogit- with a separate
>> analysis for each country.
>>
>> Think of it this way. "Fixed effects" are (unobserved) intercepts for
>> each country.  If you have enough data (you might!), you could fit a
>> separate dummy for each country, in addition to the covariate you now
>> have. This would be an "ordinary" unconditional logistic model.  This
>> model would not make sense to me.
>>
>> I am not an econometrician, so I may  completely misunderstand your
>> problem. If so, I am sure that one of the econometric experts on the
>> list will set me straight.
>>
>> -Steve
>>
>>
>> On Wed, Dec 16, 2009 at 8:25 AM, joachim jarreau
>> <joachim.jarreau@gmail.com> wrote:
>>> Hi statalist,
>>>
>>> I am trying to fit a conditional logit model on some 160,000 obs.;
>>> there are about 150 groups spanning 11 years. In the specification Id
>>> want to include 3 additional sets of fixed effects; so that the
>>> command line is of the type :
>>>
>>> clogit export x1 x2 x3  provd* secd* yeard*, group(country)
>>>
>>> where "export" is a binomial variable for positive exports in a given
>>> sector and year, from a province to a given country. "provd", "secd"
>>> and "yeard" are corresponding sets of dummies.
>>> I get "initial values not feasible" (r1400) as error message.
>>> Ive got two questions:
>>>
>>> - What is the critical aspect that makes this specification to heavy
>>> to implement/prevents convergence? what should I cut? Do you think
>>> there are too many observations, too many dummy variables, or both?
>>> (Im am trying these alternative solutions, but any indication
>>> concerning a preferred direction to search would be helpful).
>>>
>>> - In searching through precedent posts on the topic, i found a
>>> reference to a possibility to feed results from an OLS estimation into
>>> a clogit estimation. How can one do this? do you think it would help?
>>>
>>> Thank you in advance for any advice.
>>>
Steven Samuels
sjsamuels@gmail.com
18 Cantine's Island
Saugerties NY 12477
USA
845-246-0774

*
*   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