Re: st: effects-type coding of attributes in discrete choice modeling

["Nils Wloemert" <nils.wloemert@gmx.net>]

thanks a lot, this works great! 

however, I still have difficulties estimating the model. before I will estimate the nested model, I would like to replicate the non-nested logit result I got using a different software package (latent gold choice) which fits McFadden's conditional logit choice model. I computed a 1-segment solution which I assume to produce the same result as the asclogit in Stata (which also fits McFadden's model). Is this assumtion correct? Latent gold estimates the model based on choice as the dependent variable, the different attributes as independet variables and the caseID and choice-setID as indicators for the cases and alternatives.   

using clogit and linear coding of attributes, the results of both packages match quite closely. however, using effect-coding, 3 levels are omitted because of collinearity in Stata (which does not occur in latent gold) and results differ significantly between the software packages. 

using asclogit, the same 3 levels are omitted and the model does not converge. as in latent gold, choice (1,0 dummy) is defined as dependent variable, the attributes as independent variables (price linear, all other effect-coded) and respondentID serves as case identifier and alternativeID as identifier of the alternative. please note, that unlike in latent gold, it is not possible to use the choice-setID as an indicator for the alternatives in Stata, because this would mean 4 lines with the same ID per respondent (case) in the data set which is not allowed: "variable alternative has replicate levels for one or more cases; this is not allowed", i.e., a unique identifier for each alternative per respondent is used instead but the model does not converge. 

any help on how to define the model correctly is much appreciated!

thanks,

nils. 


-------- Original-Nachricht --------
> Datum: Fri, 26 Nov 2010 22:47:17 +0000 (GMT)
> Von: Shehzad Ali <drshehzad_ali@yahoo.com>
> An: statalist@hsphsun2.harvard.edu
> Betreff: Re: st: effects-type coding of attributes in discrete choice modeling

> Check xi3 and look for deviation coding.
> Shehzad
> 
> On Fri, 26 Nov 2010 04:16 Etc/GMT+12 Nils Wlömert wrote:
> 
> >Hello everyone,
> >
> >I am new to this list and fairly new to Stata, so I hope my question is
> not trivial or has been answered before. I searched the FAQs and the mailing
> list archives and I could not find anything about this specific question
> regarding choice-based conjoint, resp. discrete choice models.
> >
> >I would like to analyze data from a choice-based conjoint survey incl. 6
> attributes + a base alternative (no-choice option):
> >Attribute 1 & 2   (5 levels each) representing alternative specific price
> attributes
> >Attribute 3 - 6   (4 levels each) representing different product
> characteristics
> >+ the no-choice option
> >
> >The data set:
> >There are 8 observed choices for each respondent. Each choice is between
> 3 alternative configurations of the product and the no-choice option, i.e.,
> there are 32 lines for each respondent in the data set representing one
> alternative each (N=2540). The dependent variable (choice) is indicated by a
> 1,0 dummy variable in an extra column. There is currently one column for
> each attribute indicating the level of the attribute for the specific
> alternative (values 1-5 for the price attributes and values 1-4 for the product
> characteristics) and one column for the no-choice option (values 1 or 0).
> There are also columns indicating the respondent ID and the choice set ID.
> >
> >The analysis:
> >I have so far computed part-worths for each attribute using mlogit with
> linear (or numeric) coding of all attributes - choice being the dependent
> variable and the 6 attributes + no-choice option the independent variables.
> However, I would like to compute part-worths for every level of the product
> attributes (i.e., attributes 3 - 6 with levels 1, 2, 3, 4) using
> effects-type (or nominal) coding. Any help on how I could achieve this will be much
> appreciated!
> >
> >I would then like to compare alternative modeling approaches for the
> no-choice option with regard to the relative fit of the model (e.g., as a dummy
> variable in multinomial logit versus nested in a separate nest using
> nlogit). Also, I would like to compare different coding of the linear price
> attributes (e.g., 1,2,3,4,5 vs. mean centered -1, -0.5, 0, 0.5, 1). So any help
> on how to assign different values to attribute levels would also be much
> appreciated!
> >
> >Thank you!
> >
> >Nils.
> >*
> >*   For searches and help try:
> >*   http://www.stata.com/help.cgi?search
> >*   http://www.stata.com/support/statalist/faq
> >*   http://www.ats.ucla.edu/stat/stata/
> 
> 
> 
>       
> 
> *
> *   For searches and help try:
> *   http://www.stata.com/help.cgi?search
> *   http://www.stata.com/support/statalist/faq
> *   http://www.ats.ucla.edu/stat/stata/

-- 
GRATIS! Movie-FLAT mit über 300 Videos. 
Jetzt freischalten unter http://portal.gmx.net/de/go/maxdome
*
*   For searches and help try:
*   http://www.stata.com/help.cgi?search
*   http://www.stata.com/support/statalist/faq
*   http://www.ats.ucla.edu/stat/stata/