Altair RISE

A program to recognize and reward our most engaged community members

Nominate Yourself Now!

Conditional Logit in RM?

Hi,

I'm trying to figure out how to perform a conditional logit learner in RM.

I've seen this function in R and Stata but would love to find a way to do it in RM. (RM is my favorite!!)

It was previously suggested that I just use the Logistic Regression operator, but that doesn't produce the same results as a proper conditional logit analysis.

Form a colleague of mine:
" CL takes into account competition whereas ordinary logistic regression does not. The easiest way to verify this is to check whether the estimated probabilities sum to one in a group. If using logistic regression, there is no reason why they should because logistic regression considers every example as an entity to be classified. Contrary, CL takes a group as entity and in doing so calibrates the probabilities, so that they do sum to one across all examples in a group. Actually, the main point is not whether probabilities sum to one, but whether the estimation of probabilities considers other entries in a group."

RM can do anything, there must be a way....

Find more posts tagged with

AI Studio

Accepted answers

All comments

land

Hi,
just because I'm currious: In which cases does the Conditional Logit offer advantages over the normal logistic regression? You said something about groups, but what do these groups mean and how are examples assigned to them?

Always love to get to know new algorithms...

Greetings,
Sebastian

bobdobbs

Sebastian,

I'm far from an an expert, but from what I understand:

A CL modeler looks at cases in a group and scores them in relation to each other. Generally the probabilities will sum to one for each "group" in the data set.

Regular LR will just score every case in the data set, ignoring the group stratification.

land

Hmm,
interesting, but in which application could this be usefull? It seems to me, that the group assignment would dominate the prediction?

Greetings,
Sebastian

bobdobbs

There would be many applications.

Imagine a simple race and trying to predict the winner.

You don't care about a runner's probability vs. every race in the data set. You care about his probability vs. the other runners in this race.

eva_v

Hi,
I am wondering if anyone has found the solution to this problem, as I am also wanting to use a conditional logit (which is easily implemented in Stata or Matlab -- I can't believe RM doesn't seem to have this ability, but maybe this is my mistake).
Thanks,
Eva

Addendum: A hierarchical Bayes model would also work. I believe they are asymptotically equivalent. Unfortunately, I also can't figure out how to make a hierarchical Bayesian model....

land

Hi,
I will have to read some papers or a book about this before I can say, if we will have this feature in future. Since I'm originally a computer scientist and not a statistician, my education somehow missed this detail

If someone of you have matlab code, could recommend a paper or book, or something similar, this would be helpful. But since we are under high pressure until the release of 5.0, it will take some time until we have only estimated if we will add it...
You could always pay money to climb up the priority ladder (because open-source companys are always in need of money) but mentioning this always sounds like cheap advertisement, doesn't it? ;-)

Greetings,
Sebastian

eva_v

Hi,
Thanks for your reply!
Ken Train wrote a book on these kinds of models and all of the chapters are available for free online here: http://elsa.berkeley.edu/books/choice2.html
He also has some matlab code for it: http://elsa.berkeley.edu/~train/software.html
I hope this helps, and I look forward to seeing these kinds of models being able to be estimated in RM some day!

Eva