How to address two idenpendent random intercepts into melogit?

[Tong Liu]

I formulate a equation as y=a₀+aX+v_i+u_t The latter two term are random intercepts.
How to code it in stata? I have tried
melogit y x ||s: ||t:
But there are nested relationship between s and t and provide intercepts as u_st rather than u_t...
Coul you please help me?

Thanks!

[Erik Ruzek]

Tong,

When you say independent, you mean non-nested? If so, then look at the syntax for fitting a cross-classified model. In your case, it would be something like this:

melogit y x || _all: R.s || t:

Note that you will want to model the factor with the smallest number of units in the R. classification. Example 5 in the melogit section of Stata 15's me handbook has information on these models.

[Tong Liu]

Tong,

When you say independent, you mean non-nested? If so, then look at the syntax for fitting a cross-classified model. In your case, it would be something like this:

melogit y x || _all: R.s || t:

Note that you will want to model the factor with the smallest number of units in the R. classification. Example 5 in the melogit section of Stata 15's me handbook has information on these models.

Erik,

Thanks! If I have more than random intercepts, then I can write like
melogit y x || _all: R.s || _all:R.t: || _all: R.w ?

[Erik Ruzek]

What do you mean more than random intercepts? Are you talking about allowing the effects of predictors on the outcome to vary across groups? If so, that syntax is different.

[Tong Liu]

I Think I mean only random-intercepts..
Thanks!

[Erik Ruzek]

Yes, you can include such intercepts, but I would advise against always using the R. notation with your random intercepts, particularly if the data is partially nested. It may make the model computationally intractable. Each time you use R. with an intercept, each of the levels of the random factor is included as a separate parameter in the vector of random effects. With many random intercepts this becomes a huge design matrix. See the Marchenko article in Stata Journal in 2006 as well as Chapter 9 of Rabe-Hesketh & Skrondal's Longitudinal Modeling in Stata book for help with how to specify such models. If you do not have a partially nested model, then you may need to move to Bayesian estimation.

[Tong Liu]

Thank you Erik!