Seqlogit

[dsghosh]

Hi all
Currently I am using sequential logit of Buis (2007) to understand effect of son preference on contraceptive adoption and choice. My, main predictor is son preference, which is a gender-parity (GP) categorical variable (having 15 categories) with GP(2,1) means 2 living children with 1 living son as reference category. The response variable is contraceptive adoption and choice (1=not using, 2=using natural method, 3=using modern method). There are a number of controls. have found out the odds ratios simply using command xi: seqlogit Depvar i.Indepvars, or robust. Now I want to find out transitional probabilities of contraceptive adoption and choice (transition 1, is adoption i.e. using or not-using; and transition 2 is choice between natural and modern method ) for related to Gender-parity variable i.e. main predictor variable at each transition separately by setting other controls at their means. I can not use 'ofinterest' in the equation since it always wants 'over'. I don't have any 'over' variable. so seqlogitdecomp is also not working.

After many attempts, I just could not do it. if anybody is having syntax then pls. help me out.

[Maarten Buis]

I have two comments:

1. The decomposition performed in seqlogitdecomp only makes sense if you are willing to treat the explanatory/predictor/x-variable as continuous.
2. If that variable is continuous, you can compute the decomposition without the over() option in seqlogit if you specify the table option in seqlogitdecomp.

[dsghosh]

OK. Thanks Maarten for the comments. Does it mean that if main predictor variable is categorical (having more than 2 categories) one can not compute predicted probabilities for transition. We have to be satisfied only with odds ratio?

[Maarten Buis]

seqlogitdecomp has a very specific purpose: decompose the marginal effect of variables on the final outcome of the process modeled by seqlogit into a weighted sum of effects of this explanatory variable on different transistions and a weight for each transition. This decomposition only works when the explanatory variable is continuous.

However, there are many other things you can look at, and it seems that you are interested in something else. Maybe what you want to do is possible in such models, but that depends on what it is exactly that you want to do. For example you can compute predicted probabilities of either passing transitions or attaining a specific final outcome if predictors are categorical. You can also compute marginal effects or discrete differences on either of these probabilities using margins. The latter is usually more useful when dealing with categorical explanatory variables. You just cannot decompose these discrete differences as nicely as you can marginal effect, which is why I stated that the decomposition and seqlogitdecomp works only for continuous variables.

Here are some examples:

use "http://fmwww.bc.edu/repec/bocode/g/gss.dta", clear
                
recode degree 4=3
label define degree 0 "lt high school" ///
                    1 "high school"    ///
                    2 "junior college" ///
                    3 "college", modify
label value degree degre

seqlogit degree i.south                 ///
     c.coh##c.coh if black == 0 ,       ///
     tree(0 : 1 2 3 , 1 : 2 3 , 2 : 3 ) ///
     ofinterest(paeduc)                 ///
     levels(0=9, 1=12, 2=14, 3=16)

// In the south you expect .2 pseudo-years less education
// than in the north
margins, dydx(south)

// in the south the probability of finishing college is
// 0.4% less than in the north, but the difference is not
// significant
margins, dydx(south) predict(pr outcome(3))

// in the south the probability of entering university given
// that one finished high school is 0.1% larger than in the
// north, but the difference is not significant
margins, dydx(south) predict(trpr transition(2) choice(1))

seqlogit degree i.south                 ///
     c.coh##c.coh if black == 0 ,       ///
     tree(0 : 1 2 3 , 1 : 2 3 , 2 : 3 ) ///
     ofinterest(paeduc)                 ///
     levels(0=0, 1=0, 2=1, 3=1)
    
// In the south the probability of entering university
// (junior college and college) is 1% less than in the
// north, but the difference is not siginifcant
margins, dydx(south)

// The difference between the last two is that one is conditional
// on having passed high school and the other is not.

[dsghosh]

Thanks Maarten for the clue. I am attaching the equation, which will probably understand the problem I am facing. The equation is as follows:

xi: seqlogit rbcontrause i.nbrsexcomp i.wage i.rbnewedu i.bnewpartoccu i.bnewspaper i.btv i.bnewrelig i.V190 i.V102 i.rdchild i.bdechealth i.bwant[fw=biwt], tree(1:2 3, 2:3) or robust

In my case, rbcontrause is the response variable with 3 categories (1=not using contraception, 2=using natural method and 3=using modern method). My main predictor is nbrsexcomp, which is gender-parity composition (with 15 categories). I just want to see the marginal effect at each transition of (probability of contraceptive use at each gender-parity composition) by fixing other covariates at their means. I am writing the following syntax

.margins, dydx(nbrsexcomp)), however, it is not working. I am getting the error msg

"Warning: cannot perform check for estimable functions.
'nbrsexcomp' not found in list of covariates
r(322);
Please help.

[Maarten Buis]

I don't see why you get that error message, but I do see that even if you did not get that error message you would still not get what you want. Please see the examples I sent before.

[dsghosh]

Thanks for taking so much pain on seqlogit. I am writing the following command seeing your example:

margins, dydx(nbrsexcomp) predict(trpr transition(1) rbcontrause(1))

However, same error msg is coming "Warning: cannot perform check for estimable functions.
'nbrsexcomp' not found in list of covariates"
r(322);

I really dont understand where is the problem!!!

[Maarten Buis]

Neither do I and I cannot replicate that problem, so there is nothing I can do.

[dsghosh]

Maarten: The following command working and provide predicted probabilities of all variables:

margins, dydx(*) atmeans predict (trpr transition(1) choice(1))

margins, dydx(*) atmeans predict (trpr transition(2) choice(1))

Also, providing desired results. Thanks for your inputs and suggestions.

[Luis Ortiz]

Thanks, Saswata, for providing these two lines of command. They are useful.

I had a problem similar to yours.

Best wishes

Luis