Repeated Measures mlogit

[Simone Cheung]

Hi there, I have been reading the forum on repeated measures/panel style data and the use of mlogit. For context, my data consists of four observations per participant. Each participant watched four videos and were asked to make one of four choices for each video (1-4). I have created dummies for features of the videos - social (=1 if social setting video, =0 if sport setting video) and moms (=1 if man on man sexism and =0 if man on woman sexism).l

I have tried several options and landed with the following analysis:

1. xtlogit identify social moms socialxmoms male australian asian age age2 order where identify is a dummy (=1 if choices=2-4, 0 otherwise) and gender, ethnicity and age are the same across videos (individual characteristics)

2. mlogit speakup social moms socialxmoms male australian asian age age2 order if speakup!=1 & exclude==0, base(2) nolog vce(cl pid) where speakup = 2, 3 or 4 (whatever choice they selected)

Questions:

1. Is it right to use xtlogit in order to take into account the repeated measures component (i.e. treat it like a panel dataset but without a time component)?

2. As there is no mlogit function for panel data, is it right to use mlogit and clustard the errors around my group id? I have also tried mixlogit and femlogit. mixlogit doesn't work as it requires specific RE variables which I don't have/don't know if I have. femlogit did work but how is this different to mlogit with vce(cl pid)?

3. For both analyses above, do I progress with the usual margins, dydx(*) code to get my marginal effects?

Thank you in advance.

Cheers, Simone

[Clyde Schechter]

1. Yes. There is no need for there to be a time component with the -xt- commands unless you need to use time-series operators or autoregressive structure. Clearly neither of the latter apply in your situation, so this is fine.

2. No. Using clustered vce is a step in the right direction, but does not sufficiently account for the nesting of videos within persons. You are correct that there is no -xtmlogit- command. And neither -mixlogit- nor -femlogit- implements that model either. You can emulate multinomial logistic regression with two levels by using the -gsem- command. There is an example of this in the PDF manuals that come installed with your Stata. Run -help gsem_examples- and then scroll down the list until you find "Two-level multinomial logistic regression." (In the current version of Stata--16-- it is example 41g.) Click on the blue link and the appropriate section of the manual will open. Just adapt the code to the specifics of your data.

[Arne Risa Hole]

It is not correct that -mixlogit- cannot be used to implement a random-effects multinomial logit model - see the second (pupils’ behavior) example in https://www.stata-journal.com/articl...article=st0133 This is equivalent to the "Two-level multinomial logistic regression" which can be implemented using -gsem-.