Sequential logit model with a categorical variable of interest

[Luis Ortiz]

Dear all,

I have intended to use seqlogit for modelling educational attainment at three different levels (1=Less than university; 2=University BA; 3=University MA). These are the three categories of my dependent variable. And I know sequence logit model is a good option because these are successive transitions. Individuals who leave education before entering into university are not exposed to the possibility of just doing a BA or moving to a MA.

I have carefully explored the wonderful explanation that Maarten Buis, creator of this Stata command, made in the following link:

http://www.maartenbuis.nl/software/seqlogit.html#back3

And I have tried to reproduce the routines there with my own data.

seqlogit educa_3bis edufath female age i.imgen2 , tree( 1 : 2 3, 2 : 3) ofinterest(edufath)

Quite unfortunately, my variable of interest (edufath, father's education) is categorical. It has three categories. And I am afraid the syntax of seqlogit does not support this possibility. The variable to include in the 'ofinterest' option must be necessarily continuous. Maarten himself commented on that in the following post of this list:

https://www.statalist.org/forums/for.../2418-seqlogit

By any chance, do you know if there is any way of estimating the effect of each level of father's education for the successive transitions implicit in my dependent variable?

Many thanks for your attention

Luis Ortiz

[Phil Bromiley]

You'll increase your chances of a useful answer by following the FAQ on asking questions - provide Stata code in code delimiters, readable Stat output, and sample data using dataex.

With a user-written procedure, you probably need to read the underlying theory papers to understand what is allowable. If that doesn't help, try contacting the author.

[Luis Ortiz]

Many thanks for the guidance, Phil.

Yes, I'lll try to contact the author.

Best

Luis Ortiz

[Maarten Buis]

If you just want a sequential logit model, you can just prepare your data and estimate the model with a (set of) logit model(s). If you also want the decomposition, then there is a discussion on getting a similar decomposition for categorical variables in the article http://dx.doi.org/10.1177/0049124115591014

[Luis Ortiz]

Dear Maarten,

First of all, thanks for your contribution to Stata (seqlogit) and for your guidance.

Yes, I had read with great interest the article that you mention in your post.

I am interested in modelling a set of choices that are hierarchical and quite similar to the ones you study in your article, but I do not want to make an exercise of decomposition; I simply want to account for the fact that the probability of finding an individual in a given category is conditional on having made a prior transition.

My dependent variable is educational attainment. It is constituted by three categories: 'Less than university', 'University BA', 'University MA'. I am inclined to think that a simple multinomial logit model would not do, but perhaps I am wrong. Being in the third category is conditional upon being in the second; and being in the second is in turn conditional on scoring 0 in the first one. The situation is similar to the one you graphically represent in pg. 110 of your article, for the Dutch system of education.

This is the reason I explored in Stata... and I found seqlogit. I initially believed that seqlogit is close to what I need, but, as I said in my initial post, the 'variable of interest' (father's education) is not continuous, and it is difficult to turn into a continuous variables, since I am dealing with many different systems of education.

Referring now to your previous response, when you write about a set of logit models, I guess that you refer to the set of logit models implicit in the 'tree' option in the next line of command; am I wrong?

seqlogit educa_3bis edufath female age i.imgen2 , tree( 1 : 2 3, 2 : 3) ofinterest(edufath) 

Wouldn't it be wrong to disregard that the probability of being 'University BA' or 'University MA' is conditional on a previous choice (entering or not into university)?

Thanks for your attention

And kind regards

Luis Ortiz

[Maarten Buis]

It is the exact same model:

// open example data
sysuse nlsw88, clear

// prepare the data
gen byte edcat = cond(grade <  12, 1,     ///
                 cond(grade == 12, 2,     ///
                 cond(grade <  16, 3,4))) ///
                 if !missing(grade)
label variable edcat "respondent's education"
label define edcat 1 "< highschool"    ///
                   2 "highschool"      ///
                   3 "some college"    ///
                   4 "college"            
label value edcat edcat

// method 1
seqlogit edcat i.south i.race, tree(1:2 3 4, 2 : 3 4 , 3: 4)

// method 2
gen hs = edcat >= 2 if edcat < .
gen sc = edcat >= 3 if edcat >= 2 & edcat < .
gen c  = edcat == 4 if edcat >= 3 & edcat < .

logit hs i.south i.race
logit sc i.south i.race
logit c  i.south i.race

In fact, what happens inside seqlogit is that these separate (m)logits are used to create starting values, and you see that seqlogit converges in 1 iteration.

[Luis Ortiz]

Thanks, Maarten.....¡

Very useful

Best regards

Luis Ortiz