sequential logit

[Deniz Guvercin]

I am running sequential logistic regression to infer about usage of drug on crime. there are 4 waves of interviews. I am trying to construct tree at which I am totally confused.
First I define my two transitions as 0000, and 0111, and then run seqlogit for the tree (1:2) which didnt give desired results (coef<0). I am just wondering if I am making mistakes at the stage of constructing tree


g drugtrans =1 if ayanydrug==0 & byanydrug==0 & cyanydrug==0 & dyanydrug==0 (948 missing values generated) .
replace drugtrans =2 if ayanydrug==0 & byanydrug==1 & cyanydrug==1 & dyanydrug==1 (60 real changes made) .

seqlogit drugtrans i.respsex i.arst, tree (1:2)
Transition tree:
Transition 1: 1 : 2
Computing starting values for:
Transition 1
Iteration 0: log likelihood = -249.33364
Iteration 1: log likelihood = -249.33364
	Number of obs =	1,649
LR chi2(3) =	16.75
Log likelihood = -249.33364	Prob > chi2 =	0.0008
drugtrans Coef. Std. Err. z	P>z [95% Conf.	Interval]
respsex
female .1750619 .2702427 0.65	0.517 -.3546041	.7047279
arst
2 -.5273682 1.163988 -0.45	0.650 -2.808742	1.754006
3 -2.129618 .4539332 -4.69	0.000 -3.019311	-1.239925
_cons -1.367098 .4349749 -3.14	0.002 -2.219633	-.5145628

[Maarten Buis]

The output is unreadable, can you repost it using the [CODE] deliminators (the button with the "#")?

[Deniz Guvercin]

 g drugtrans =1 if ayanydrug==0 & byanydrug==0 & cyanydrug==0 & dyanydrug==0
(948 missing values generated)

. replace drugtrans =2 if ayanydrug==0 & byanydrug==1 & cyanydrug==1 & dyanydrug==1
(60 real changes made)

. seqlogit drugtrans i.respsex i.arst, tree (1:2)

Transition tree:

Transition 1: 1 : 2

Computing starting values for: 

Transition 1

Iteration 0:   log likelihood = -249.33364  
Iteration 1:   log likelihood = -249.33364  

Number of obs     =      1,649
LR chi2(3)        =      16.75
Log likelihood = -249.33364                     Prob > chi2       =     0.0008


drugtrans       Coef.   Std. Err.      z    P>z     [95% Conf. Interval]

respsex 
female     .1750619   .2702427     0.65   0.517    -.3546041    .7047279

arst 
2    -.5273682   1.163988    -0.45   0.650    -2.808742    1.754006
3    -2.129618   .4539332    -4.69   0.000    -3.019311   -1.239925

_cons   -1.367098   .4349749    -3.14   0.002    -2.219633   -.5145628

[Maarten Buis]

For this you don't need seqlogit, as there is no sequence. You can just do this with regular logit.

I don't see why negative coefficients pose a problem. Maybe you are expecting odds ratios? To get odds ratios you need to specify the or option just as in regular logit.

[Deniz Guvercin]

Then, I will define different sequences such as 0000 to 1111; 0000 to 1010 ; 1000 to 0000, and then run regular logit and interpret each transition seperately. Thank you for your guidance.

[Maarten Buis]

You are being very cryptic. Remember, we know nothing about your research, data, variable, coding of your variables, etc. I suspect that seqlogit is not right for your problem, but it is impossible for me to be sure given the lack of information you have given us.

[Deniz Guvercin]

ayanydrug is 0 if the respondent did not use drug in last year and & byanydrug, cyanydrug, dyanydrug==0 if the respondent did not use the drug in second third and fourth year.
Arrst is 0 if the respondent didnt get arrested last year. So if the transition is from 0000 to 0111 the drug gets habituated. I am trying to understand how beneficial is crime income,education (for each wave), etc. for this transition to occur? How about if transition is from 0000 to 1111 or from 1111 to 0000?
Is 0000 to 1111 the transition or we compare two different complete transitions here?
Thank you

[Maarten Buis]

You are not looking at transitions, but comparing complete trajectories. A transition is a step from one wave to the next, you are looking at all 4 waves together.

Based on your questions I suspect that you haven't finished thinking about the exact aims of your research project and how your data can (or cannot) help answer those. This is normal: when I start a new project I typically think I have a clear idea of what I want to do, then I try start preparing the data and planning my analysis and I find out that it is all much more complicated than that. (This is a variation of: no plan survives contact with the enemy) Nowadays I expect that, and treat finding the research question, developing theory, and doing the analysis as an iterative process rather than a sequential process.

[Deniz Guvercin]

I think comparing trajectories also is valuable. I think setting up seperate logits for each transition might help to derive some conclusions?
I actually totally agree on not completely comprehending the (prior) intuition in the modelling framework where different intuitions blur the prior ones and completely put everything in disorder.
Thank you for your time!