Re: st: Convergence never achieved with MI impute chained

[Maarten Buis <maartenlbuis@gmail.com>]

I have just seen that you used only the first digit of the ISIC
classification. However, it contains lots of sparse categories(*).
These will cause problems. Also inspect education for sparse
categories. You'll need to combine those sparse categories with
"adjacent" categories in order to get sufficiently filled cells. Also
look at a cross tabulation of industry and education, and see if there
aren't any cells that are too empty. That will probably mean a second
round of merging categories.

I would not use ordered models or mvn for imputing industry, that just
does not make sense.

Hope this helps,
Maarten

(*) If this is recent data from a western country than you have made a
coding error. In that case there are way way way too many farmers.

On Thu, Jun 21, 2012 at 1:46 PM, Lena Lindbjerg Sperling
<lenalindbjergsperling@gmail.com> wrote:
>>
>> Thank you for your answer!
>>
>> It does seem though that all occupations are represented in both private and public sectors.
>> And I also have another data set where I only impute educational level, industry (ISIC 3 category) and wage and I still get not convergence, even though that's just one mlogit, one ologit and one pmm...so that doesn't seem to be the problem.
>>
>> I got a result out for the mi xeq 0: mlogit for industry however and it looks like this:
>> ->    mlogit  industry
> Iteration       0:00    log     likelihood      =       -4875.9554
> Iteration       1:00    log     likelihood      =       -4875.9554
> Multinomial     logistic        regression      Number  of      obs     =
> LR      chi2(0) =       0
> Prob    >       chi2    =       .
> Log     likelihood      =       -4875.9554      Pseudo  R2      =
> industry        Coef.   Std.    Err.    z       P>z     [95%
> Agriculture__Hunting__etc_      (base   outcome)
> Mining
> _cons   -4.982464       0.2896632       -17.2   0       -5.550194       -4.414735
> Manufacturing
> _cons   -2.671581       0.0939994       -28.42  0       -2.855816       -2.487345
> Public_services
> _cons   -3.42432        0.134593        -25.44  0       -3.688117       -3.160522
> Construction
> _cons   -3.204691       0.1210617       -26.47  0       -3.441968       -2.967415
> Retail__Hotels
> _cons   -1.714798       0.0612048       -28.02  0       -1.834758       -1.594839
> Transport_and_telecomnunications
> _cons   -4.759321       0.2593031       -18.35  0       -5.267546       -4.251096
> Finance_and_business_serv_
> _cons   -6.368759       0.5778449       -11.02  0       -7.501314       -5.236204
> Communal_services
> _cons   -0.830113       0.0433825       -19.13  0       -0.9151412      -0.7450848
> Others_not_well_specified
> _cons   -1.753638       0.0622235       -28.18  0       -1.875594       -1.631683
>>
>> Should I use something else to impute this? It runs from 1 to 10 so maybe ordered is better? I get convergence if I use ordered logit for industry and occupation. They really shouldn't be ordered, but how important is that choice?
>>
>>
>> I can get results out if I use mvn, but is that a very bad idea? Seems like the literature disagrees quite a bit on how severe it is to assume normality?
>>
>> Best,
>> Lena
>>
>> Den Jun 21, 2012 kl. 10:48 AM skrev Maarten Buis:
>>
>>> On Thu, Jun 21, 2012 at 10:15 AM, Lena Lindbjerg Sperling wrote:
>>>> I just looked at the mail again, and the data is not as bad as it looks, as I'm only imputing on the employed population (lstatus==1) and when we only look at them mi describe shows:
>>>> mi describe
>>>>
>>>> Style:  wide
>>>>         last mi update 21jun2012 10:03:51, 18 seconds ago
>>>>
>>>> Obs.:   complete        2,702
>>>>         incomplete        912  (M = 0 imputations)
>>>>         ---------------------
>>>>         total           3,614
>>>>
>>>> Vars.:  imputed:  7; occup(126) ocusec(144) whours(167) edulevel(171) ocu(228) industry(204) mwage(598)
>>>
>>> Just looking at the variable names I suspect that this is an extremely
>>> hard model to estimate. How many categories do the variables occup,
>>> ocusec, ocu, and industry have? Are there combinations of three or
>>> less of these that for some observations perfectly predict one or more
>>> remaining variables? For example, if we know that someone is a mayor
>>> than we also know that (s)he is working in the public sector.
>>>
>>> <snip>
>>>> Iteration 14:  log pseudolikelihood = -2454486.7  (not concave)
>>>> Not completely sure what this means. Can you see where things are wrong from this?
>>>
>>> It means that this sub-model did not converge, probably because of the
>>> problems indicated above.
>>>
>>>> When I use -mi xeq 0: mlogit - the result is:
>>>> m=0 data:
>>>> -> mlogit
>>>> last estimates not found
>>>> r(301);
>>>>
>>>> But I thought it was the observed data...which should be there?
>>>
>>> What you asked for was for Stata to replay the last -mlogit- command,
>>> and it replied that the last command wasn't -mlogit-. You probably
>>> pressed break before the model finished estimating, which makes sense
>>> if it did not converge.
>>>
>>> Hope this helps,
>>> Maarten
>>>
>>> --------------------------
>>> Maarten L. Buis
>>> Institut fuer Soziologie
>>> Universitaet Tuebingen
>>> Wilhelmstrasse 36
>>> 72074 Tuebingen
>>> Germany
>>>
>>>
>>> http://www.maartenbuis.nl
>>> --------------------------
>>>
>>> *
>>> *   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/
>>
>
>
> *
> *   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/



-- 
--------------------------
Maarten L. Buis
Institut fuer Soziologie
Universitaet Tuebingen
Wilhelmstrasse 36
72074 Tuebingen
Germany


http://www.maartenbuis.nl
--------------------------

*
*   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/