[R] reduced set of alternatives in package mlogit

Jose Marcos Ferraro jose.ferraro at LOGITeng.com
Wed Apr 13 19:18:35 CEST 2016

code? example data?  We can only guess based on your vague post.

"PLEASE do read the posting guide http://www.R-project.org/posting-guide.html
and provide commented, minimal, self-contained, reproducible code."

Moreover, this sounds like a statistical question, not a question about R programming, and so might be more appropriate for a statistical list like stats.stackexchange.com<http://stats.stackexchange.com>  .


Bert Gunter

Sorry if I was not clear enough, but  there is hardly any code to show.
The problem is that a parameter or function is lacking (or , mostly likely, I can't find it), so in some sense the problem itself is that  there is no code to show.

In what follows choice situations , alternatives, wide, and variables have the same meaning that they have on the mlogit documentation. All variables are alternative specific.

1)I want to estimate a multinomial Logit  using the mlogit package

2)I have a dataset, made of choice situations

3)There is a set of alternatives

4)in some choice situations, not all alternatives were available, but only a subset of them. So there are no variables for the unavailable alternatives and the chosen alternative evidently  belongs to the set of available ones.

5)I use mlogit.data to prepare the dataset from a "wide" dataframe . There is no option to have only a subset of alternatives and the resulting object will have them all , that is, there will be a line for every alternative and every choice situation, even if in reality some of them were not available. The variables of these alternatives did not exist, so must be filled with 0s or any other made up value

6) If ones estimate a model from this data it will be wrong

7) It is possible to get an "almost right" model by using a dummy variable marking which alternatives are unavailable, for as it is only used in alternatives that are never chosen, its coefficient will get negative with big absolute value, in practice giving almost 0% probability for them

8)this is a workaround because it obligates the model to estimate a number that should be -infinity and this is known in advance, so it's ugly and difficult to know what the numeric consequences are as the coefficient can never converge. In fact, I don't use it the way I described for these reasons, preferring a more complex but almost equivalent formulation. The important point is that I want a clean solution, not a workaround

9)I demand simply if mlogit package has such functionality

I meant that in some choice situations there are some alternatives missing, but the available alternatives are known to everybody(both the one that made the choice as well as to who collected the data).
For future reference, I would like to post here that I found the answer.
Apparently it is not possible if one uses mlogit.data  with shape = “wide”, but it is if one uses it with shape = “long” .
So basically one can create an alternative specific variable with availability (let’s call it is_avaliable) and use mlogit.data  normally that is :
all_avaliable <- mlogit.data(df , shape = “wide” , …)
then one can subset it
real_avaliability <-  all_avaliable[all_avaliable$is_avaliable ,]
and resend it through mlogit.data with format long
mlogit.data(real_avaliability , shape = "long" , alt.var = "alt" ,  chid.var = "chid", …)
please observe that alt and chid will have been created by the first call to mlogit.data

	[[alternative HTML version deleted]]

More information about the R-help mailing list