Notice: On April 23, 2014, Statalist moved from an email list to a forum, based at statalist.org.
[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]
Re: st: mlogtest after mlogit
From
Chiara Mussida <[email protected]>
To
[email protected]
Subject
Re: st: mlogtest after mlogit
Date
Tue, 25 Oct 2011 13:55:07 +0200
On 25 October 2011 13:30, Nick Cox <[email protected]> wrote:
> Now we are really getting somewhere. "__000000" is a temporary name;
> in essence it surely does not belong there. I guess now that you have
> tickled a bug in one of the -spost- routines and, as recommended
> already in this thread, you should bring this to the attention of the
> authors, who as said are not active members of Statalist.
>
> Note that -_pecats- does find all your categories 1/9 which supports
> my earlier wild conjecture that their frequencies were not an issue.
>
> So, the warning message itself now appears to be an error, but there
> is no obvious reason to credit the rest of the output if -mlogtest- is
> misunderstanding what is going on with your model and data. You did
> not comment on the fact that -r(refval) is missing in your output, so
> other difficulties may lie undetected.
>
> Nick
>
Surely, I'm going to bring the attention of my problem to the authors
of mlogtest ado.
Nonetheless, the estimates of the mlogit for 9 categories do not seem
to be biased:
the results are consistent of previous findings obtained by running 3 separate
mlogit (3 mlogit with common origin state). My attempt to put all the
categories/outcome
together is to avoid to get mixed up with sample selection issues.
> On Tue, Oct 25, 2011 at 12:18 PM, Chiara Mussida <[email protected]> wrote:
>> On 25 October 2011 12:57, Nick Cox <[email protected]> wrote:
>>> If you look at the code of -mlogtest- the warning message that worries
>>> you is when a helper program -_pecats- produces output that is
>>> inconsistent. After your -mlogit- call you can run -_pecats- followed
>>> by -return list-. In a well-behaved case, you will get output like
>>> this in which r(numcats) matches the number of elements in r(catvals).
>>> Your output will probably not match up. I don't know why that might
>>> be, but I doubt that the category frequencies are somehow the issue.
>>> (For "STATA" read "Stata".)
>>>
>>> . _pecats
>>>
>>> . return list
>>>
>>> scalars:
>>> r(numcats) = 9
>>> r(refval) = 3
>>>
>>> macros:
>>> r(catnms8) : "1 2 4 5 6 7 8 9 3"
>>> r(catvals) : "1 2 4 5 6 7 8 9 3"
>>> r(catnms) : "1 2 4 5 6 7 8 9 3"
>>> r(refnm) : "3"
>>>
>> Exactly Nick and thanks: when I type -_pecats- and -return list- I get:
>>
>> return list
>>
>> scalars:
>> r(numcats) = 9
>> r(refval) = .
>>
>> macros:
>> r(catnms8) : "1 2 3 4 5 6 7 8 9"
>> r(catvals) : "1 2 3 4 5 6 7 8 9 __000000"
>> r(catnms) : "1 2 3 4 5 6 7 8 9"
>>
>> where as you correctly assumed, I do not have the match between
>> r(numcats) and r(catvals), but something like __000000 which does not
>> allow a perfect matching. I still do not know why it happens.
>>
>>
>>
>>
>>
>>>
>>> On Tue, Oct 25, 2011 at 11:10 AM, Chiara Mussida <[email protected]> wrote:
>>>> On 25 October 2011 01:10, Richard Williams
>>>> <[email protected]> wrote:
>>>>> At 05:59 AM 10/24/2011, Muhammad Anees wrote:
>>>>>>
>>>>>> have you tried -mlogtest, all- to verify if these categories have no
>>>>>> other issues. Otherwise the test on -combine- might have resulted
>>>>>> becaure some of the categories had small or no observations so
>>>>>> checking the condition of equal coefficients from -comibe- did not
>>>>>> work.
>>>>>
>>>>> I agree - I get nervous about using multiple-outcome commands like mlogit
>>>>> with lots and lots of independent variables. You may be spreading the data
>>>>> too thin. But, before taking this too much further, I'd like the original
>>>>> poster to confirm that the most current version of mlogtest is indeed being
>>>>> used. Otherwise we might be talking about a problem that was fixed 6 months
>>>>> ago. Also, it might be good to present a frequency of the dependent
>>>>> variable. Long and Freese's commands are sometimes pickier about coding than
>>>>> Stata is, e.g. they sometimes don't like non-integer coding. Also, you would
>>>>> see if some of the categories have very small frequency counts. Finally, I
>>>>> would run a simple model with only one or two independent variables followed
>>>>> by mlogtest. If the simple model works and the more complicated one doesn't,
>>>>> that might indicate problems with one or more of the added variables or with
>>>>> the data being spread too thin to do the test.
>>>>>
>>>>>
>>>> Dear All, I confirm that my current version of mlogtest is and was the
>>>> one indicated by Nick, precisely:
>>>>
>>>> . which mlogtest, all
>>>>
>>>> C:\Program Files\Stata12\ado\updates\m\mlogtest.ado
>>>> *! version 1.7.6 jsl 2009-10-18
>>>>
>>>> in terms of model estimates, I guess that the issue is related to the
>>>> relative frequency of my dependent variables categories':
>>>>
>>>> ta transition
>>>>
>>>> transition | Freq. Percent Cum.
>>>> ------------+-----------------------------------
>>>> 1 | 271 0.70 0.70
>>>> 2 | 132 0.34 1.04
>>>> 3 | 1,119 2.90 3.94
>>>> 4 | 379 0.98 4.93
>>>> 5 | 722 1.87 6.80
>>>> 6 | 13,959 36.17 42.97
>>>> 7 | 388 1.01 43.98
>>>> 8 | 168 0.44 44.41
>>>> 9 | 21,450 55.59 100.00
>>>> ------------+-----------------------------------
>>>> Total | 38,588 100.00
>>>>
>>>> e.g., categories 2 and 8 might be too small. Now, I cannot collapse my
>>>> dep variable in a reduced number of categories and I hope that
>>>> notwithstanding the STATA alert message after typing the test command
>>>> (below I copied all the results) do not bias my results. ps: for the
>>>> test N=25441 since the model estimates are referred to a subsample
>>>> (aged 15-64) of the overall population (38588).
>>>>
>>>> mlogtest, c
>>>>
>>>> Problem determining number of categories.
>>>>
>>>> **** Wald tests for combining alternatives (N=25441)
>>>>
>>>> Ho: All coefficients except intercepts associated with a given pair
>>>> of alternatives are 0 (i.e., alternatives can be combined).
>>>>
>>>> Alternatives tested| chi2 df P>chi2
>>>> -------------------+------------------------
>>>> 1- 2 | 18.576 5 0.002
>>>> 1- 3 | 5.990 5 0.307
>>>> 1- 4 | 13.565 5 0.019
>>>> 1- 5 | 148.448 5 0.000
>>>> 1- 6 | 178.434 5 0.000
>>>> 1- 7 | 33.226 5 0.000
>>>> 1- 8 | 68.938 5 0.000
>>>> 1- 9 | 311.133 5 0.000
>>>> 2- 3 | 19.911 5 0.001
>>>> 2- 4 | 23.931 5 0.000
>>>> 2- 5 | 68.619 5 0.000
>>>> 2- 6 | 68.197 5 0.000
>>>> 2- 7 | 24.027 5 0.000
>>>> 2- 8 | 55.161 5 0.000
>>>> 2- 9 | 127.081 5 0.000
>>>> 3- 4 | 36.426 5 0.000
>>>> 3- 5 | 438.391 5 0.000
>>>> 3- 6 | 703.923 5 0.000
>>>> 3- 7 | 59.553 5 0.000
>>>> 3- 8 | 103.537 5 0.000
>>>> 3- 9 | 1130.422 5 0.000
>>>> 4- 5 | 197.002 5 0.000
>>>> 4- 6 | 103.387 5 0.000
>>>> 4- 7 | 91.631 5 0.000
>>>> 4- 8 | 132.381 5 0.000
>>>> 4- 9 | 530.943 5 0.000
>>>> 5- 6 | 592.783 5 0.000
>>>> 5- 7 | 192.566 5 0.000
>>>> 5- 8 | 142.185 5 0.000
>>>> 5- 9 | 281.162 5 0.000
>>>> 6- 7 | 520.969 5 0.000
>>>> 6- 8 | 430.911 5 0.000
>>>> 6- 9 | 5946.722 5 0.000
>>>> 7- 8 | 24.867 5 0.000
>>>> 7- 9 | 231.357 5 0.000
>>>> 8- 9 | 82.933 5 0.000
>>>> --------------------------------------------
>>>>
>
> *
> * 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/
>
--
Chiara Mussida
PhD candidate
Doctoral school of Economic Policy
Catholic University, Piacenza (Italy)
*
* 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/