Checking in Manual V6 (examples 6.4 and 6.15), I realise that modeling @ multiple time points (=5) for categorical (=NOMINAL) dependent variable modeling CANNOT be done in MPlus. The modeling I refer to is an extension to classic discrete choice models (McFadden) with the same choice tasks repeated over several discrete time points. My current model design entails 5 discrete choices + no option. Independent variables forming the stimuli: 1 categorical attribute (5 nominal levels [food labels] out of which 2 can also be nested), 1 categorical attribute (food categories treatable as nominal), 1 ordered categorical attribute (2 levels). Please confirm.
Modeling of nominal outcomes at multiple time points can be done in Mplus. One approach is latent transition modeling (hidden Markov modeling). See the User's Guide for examples - although none may show a nominal outcome, this is possible.
Other models to correlate the outcomes over time are also possible.
Looked at 8.13 and 8.14 which refer to the suggestion above. Still face problems linking discrete choice modeling (DCM) and latent transition modeling (LTM); apologies-the problem is mine.
The data #usually# are in a form of rows per respondent for each choice tasks (here=5)- each row reflecting attribute level combinations X No of tasks person is exposed to. The DV column in each row is 0/1 (0=if not; 1=if selected).
The issue is: I am interested in computing the #utilities# of the independent attributes (read above) as revealed through their choices. This does not look the case in example 8.13 where c1, c2 reflect latent classes of unobserved heterogeneity.
b) Also another issue.. If I control for the impact of latent covariates (CFA factors with items), how do I do that? Element x in example 8.13 acts as a covariate, but can we have CFA factors instead? One would then estimate the influence of factors upon both the choices #per se# (=NOMINAL outcomes) -observed choices or latent factors of these choices AND their influence as moderators upon the attribute --> outcome impact.
These in a dynamic mode.
Thank you for clarifying on these two distinct issues.
Discrete-choice modeling involves utilities as you say and as far as I remember these are continuous latent variables underlying each of the different categories of a nominal outcome (where the outcome is chosen that has the highest l.v. score). That's the part I don't think Mplus can model yet, but I may be wrong. Any input from others?
b) Yes, you can have a CFA model and predict from those factors, etc.
What is possible in Mplus is to have a 2-level model for nominal outcomes. That is, for example, a random intercept model. For your applicaton, level 1 would represent variation across time and level 2 variation across individuals. This is one way of accounting for the longitudinal correlation. But it isn't discrete-choice modeling.
Has anyone followd up on this. Theer are many who pretend to model discrete-choice models in Mplus (including Temme, mentioned in another thread), but so far I was not able to replicate a standard McFadden MNL discrete-choice model in Mplus. Thanks for any input or code...
I am not an expert on discrete-choice modeling, but isn't it a case of the multinomial logit model? That's the way it is described on pages 28-29 of the 2004 Chapman-Hall latent variable book by Skrondal and Rabe-Hesketh.
David James posted on Monday, December 14, 2015 - 8:55 pm
Hi Mplus community,
I was also wondering if anyone has an example of completing an MNL model for Discrete choice modelling?
I would be very interested in reviewing an example
Temme, D., Paulssen, M., & Dannewald, T. (2008). Incorporating latent variables into discrete choice models – A simultaneous estimation approach using SEM software. BuR – Business Research, 1, 220-237. download paper contact first author show abstract
David James posted on Wednesday, December 16, 2015 - 5:40 pm
I was hoping to review another example from others that have completed MNL modelling.
David James posted on Wednesday, December 16, 2015 - 6:57 pm
Thank you for the assistance.
My main question is around why linear and equality constraints were used in the temme paper?
Multinomial logit model choice#1 ON tt_pt_pto (1); !equality constraint choice#1 ON tt_c_co (p2); !linear constraint choice#1 ON dist_bus dist_opt; choice#1 ON age gender income; choice#1 ON flex cc safe; choice#2 ON tt_pt_cpt (1); !equality constraint choice#2 ON tt_c_co (p2); !linear constraint choice#2 ON tt_c_cpt (p1); !linear constraint choice#2 ON dist_bus dist_opt; choice#2 ON age gender income; choice#2 ON flex cc safe; [choice#1 choice#2]; !alternative-specific constants
The Author does NOT respond and I suspect the MPLUS model is flawed. Indeed the reason to pose some constraint is due to the fact that choice data are different from usual mlogit data. For example, in Stata you have the Mlogit and the ASClogit command, and discrete choice si modelled with ASClogit which is not available in Mplus. However, I could not reproduce Stata (or NLOGIT results) in Mplus using Temme's approach.