Re: Re: Re: st: update seqlogit

[angelrlaso@gmail.com]

Dear Maarten,

You were right: there were some empty cells in the table outcome by origin  
by sex (in the death row), so I've dropped this outcome. Now the categories  
are: 1 moved 2 absent 3 long absence 4 refusal 5 interviewed.

In relation with the "multinomial" transition, I definitively do not want  
to merge the categories absent (what mainly means "at work") and long  
absence (mainly meaning "on vacation"). But if I follow your second  
proposal, I don't understand very well what I'm modelling. It seems that  
the transition 3 4: 5 6 (that now should be 2 3: 4 5) is just merging the  
two categories and the transition 3:4 (now 2:3) is the transition between  
being absent and being in long absence, which for me it is not a transition  
(you are either absent or in long absence, but both are deadways in  
comparison with being at home).


One other thing:

 From bivariate results, I know that having moved is more frequent among  
males, until 3.4 decades of age and among non-Spaniards (what makes sense  
in a door-step survey).

This is supported in a logistic analysis of having moved again the rest of  
the categories.

(The codes for variables are:
sexo 0 female 1 male
_Igeograf2_1 being Latin-american vs being Spanish
_Igeograf2_2 being Eastern European vs being Spanish
_Igeograf2_3 coming from othe medium-low income countries vs being Spanish
_Igeograf2_4 coming from high income countries vs being Spanish
age in decades (edtd) is broken at 3.4 and 6.4)

. xi: logistic moved edtd1 edtd2 edtd3 rentmil i.geograf2*sexo, cluster(psu)

Logistic regression Number of obs = 11884
Wald chi2(13) = 1008,74
Prob > chi2 = 0,0000
Log pseudolikelihood = -4858,2263 Pseudo R2 = 0,1076

(Std. Err. adjusted for 1209 clusters in psu)

Robust
moved Odds Ratio Std. Err. z P>z ]

sexo 1,106467 ,0653453 1,71 0,087
edtd1 1,486104 ,0944023 6,24 0,000
edtd2 ,5813945 ,0205634 -15,33 0,000
edtd3 1,720185 ,1036952 9,00 0,000
rentmil 1,004475 ,0053208 0,84 0,399
_Igeograf2_1 4,299125 ,5710822 10,98 0,000
_Igeograf2_2 7,14223 1,115184 12,59 0,000
_Igeograf2_3 4,381366 1,041568 6,21 0,000
_Igeograf2_4 2,632966 ,6537054 3,90 0,000
_IgeoXsexo_1 1,016513 ,1917234 0,09 0,931
_IgeoXsexo_2 ,9295757 ,2048988 -0,33 0,740
_IgeoXsexo_3 1,251563 ,3775649 0,74 0,457
_IgeoXsexo_4 2,251334 ,7256011 2,52 0,012


Seqlogit coefficients are congruent with this only for the sex variable,  
but not for age nor origin

. xi: seqlogit incseqd sexo edtd1 edtd2 edtd3 rentmil i.geograf2, tree(1: 2  
3 4 5 , 2 3: 4 5 , 2 :3 , 4 : 5) ///
> ofinterest (sexo) over(i.geograf2 edtd1 edtd2 edtd3) or cluster(psu)

Robust
incseqd Odds Ratio Std. Err. z P>z ]

_2_3_4_5v1
sexo 1,823968 ,6438496 1,70 0,089
edtd1 ,762823 ,0684919 -3,02 0,003
edtd2 1,748463 ,0904007 10,81 0,000
edtd3 ,5161561 ,0385954 -8,84 0,000
rentmil ,9955007 ,0052792 -0,85 0,395
_Igeograf2_1 ,2299126 ,0310905 -10,87 0,000
_Igeograf2_2 ,1397021 ,022119 -12,43 0,000
_Igeograf2_3 ,2258214 ,053916 -6,23 0,000
_Igeograf2_4 ,3729059 ,0934941 -3,93 0,000
_sexo_X__I~1 1,003078 ,1932901 0,02 0,987
_sexo_X__I~2 1,085557 ,2416302 0,37 0,712
_sexo_X__I~3 ,8142282 ,2481748 -0,67 0,500
_sexo_X__I~4 ,4677205 ,1517807 -2,34 0,019
_sexo_X_ed~1 ,7895073 ,0960845 -1,94 0,052
_sexo_X_ed~2 ,9480818 ,0663292 -0,76 0,446
_sexo_X_ed~3 1,479669 ,1970169 2,94 0,003


Have I specified something wrong in seqlogit?

Many thanks,

Angel Rodriguez-Laso


El 17/02/2010 16:22, Maarten buis <maartenbuis@yahoo.co.uk> escribió:


> --- On Wed, 17/2/10, angelrlaso@gmail.com wrote:

> > I'm having trouble with the definition of one of the

> > transitions.

> >

> > The dependent variable (incseq) codes are: 1 Dead 2 Has

> > moved 3 Absent 4 Long absence 5 Refusal 6 Interviewed.

> >

> > My transitions are:

> >

> > 1: Dead: rest of the possibilites

> > 2: Has moved: absent long-absence refusal interviewed

> > 3: Absent or long absence: refusal interviewed

> > 4: Refusal: Interviewed



> You either need a fifth transition "absent : long absence"

> or you need to combine the absent and long absence

> categories. So in terms of Stata:

> For the first solution use the option

> tree(1: 2 3 4 5 6, 2: 3 4 5 6, 3 4 : 5 6, 3 : 4, 5 : 6)



> For the second solution first type:



> gen byte incseq2 = incseq

> recode incseq2 (4 = 3)



> than use -seqlogit- with incseq2 as dependent variable

> and the following option:

> tree(1: 2 3 5 6, 2: 3 5 6, 3 : 5 6, 5 : 6)

> > I've tried then:

> >

> > xi: seqlogit

> > incseq sexo edtd1 edtd2 edtd3 rentmil i.geograf2, tree(1:2 3

> > 4 5 6, 2:3 4 5 6, 3: 5 6, 4: 5 6, 5:6) ///

> > ofinterest

> > (sexo) over(i.geograf2 edtd1 edtd2 edtd3) or

> > cluster(psu)

> >

> > Stata is now in

> >

> > Iteration 72: log pseudolikelihood = -16765,835 (not

> > concave)



> This shouldn't happen, the starting values are exactly correct,

> so it should converge in the first itteration. There is probably

> some form of perfect colinearity or perfect prediction going on.

> That can easily happen with -xi-. Do for instance all transitions

> happen in all categories of geograf2?



> 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/