It is my understanding that they ARE using a binary model. In fact, they even discuss "exploding" the model to count second and third place finishers as "winners". Otherwise, how can one calculate the probability of the positive class (winner)? If I'm mistaken and they are in fact predicting rank, would you please show me where that is in their paper.
Thanks! -N On 8/31/09 7:17 PM, Achim Zeileis wrote: > On Mon, 31 Aug 2009, Noah Silverman wrote: > >> Um..... I did my research. Have been for years. I assume you're >> referring to Boltman and Chapmanm "A multinomial logit model for >> handicapping horse races" included in "Efficiency of racetrack >> betting markets". Page 155 references what they call a "multinomial >> model". From equation 14 in their paper, it appears as if they are >> calculating "Utility" of a horse as a number. Far from what I >> understand a traditional "Multinomial" model is. >> >> The seminar that I referenced discussed using a probit model instead >> of a logit model. Since the Boltman and Chapman application didn't >> really have multiple discreet choices, I'm not sure how the probit >> model would. Hence my inquiry. > > But of course it has multiple choices (finishing ranks instead of > binary win/lose), otherwise it wouldn't be very multinomial, would it? > Z > >> >> >> On 8/31/09 6:23 PM, Achim Zeileis wrote: >>> On Mon, 31 Aug 2009, Noah Silverman wrote: >>> >>>> I get that. >>>> >>>> Still trying to figure out what the "multi" nominal labels they >>>> used were. That's why I passed on the reference to the seminar >>>> summary. >>> >>> So that I could do the research for you? Come on...the usual >>> strategy applies: Look at the references! (Hint: The information is >>> in the Bolton and Chapman paper.) >>> Z >>> >>>> >>>> On 8/31/09 5:40 PM, Achim Zeileis wrote: >>>>> On Mon, 31 Aug 2009, Noah Silverman wrote: >>>>> >>>>>> Thanks Achim, >>>>>> >>>>>> I discovered the Journal article just after posting this >>>>>> question. It did help explain more. >>>>>> >>>>>> My original inspiration for looking at this package came from a >>>>>> seminar "summary" given in 2002. Unfortunately , I can not find >>>>>> any actual published paper or lecture notes that explain the >>>>>> lecturer's application of the MNP. >>>>>> >>>>>> Here is a link to the PDF of the summary: >>>>>> http://www-stat.stanford.edu/seminars/stat/abstracts2001-2002/gu.pdf >>>>>> >>>>>> Most of the other published research on using logit or probit >>>>>> models for horseracing data use a binary label of win/lose. So, >>>>>> my thought was that they were using the same for this application. >>>>>> >>>>>> Any thoughts? >>>>> >>>>> As I said in my last mail: *Multi*nomial probit typically conveys >>>>> more than 2 choices while *bi*nomial probit conveys exactly 2 >>>>> choices. >>>>> Z >>>>> >>>>>> -- >>>>>> Noah >>>>>> >>>>>> >>>>>> On 8/31/09 5:07 PM, Achim Zeileis wrote: >>>>>>> On Mon, 31 Aug 2009, Noah Silverman wrote: >>>>>>> >>>>>>>> Hello, >>>>>>>> >>>>>>>> I want to start testing using the MNP probit function in stead >>>>>>>> of the lrm function in my current experiment. >>>>>>>> >>>>>>>> I have one dependant label and two independent varaibles. >>>>>>>> >>>>>>>> The lrm is simple >>>>>>>> >>>>>>>> model <- lrm(label ~ val1 + val2) >>>>>>>> >>>>>>>> I tried the same thing with the mnp function and got an error >>>>>>>> that I don't understand >>>>>>>> >>>>>>>> model <- mnp(label ~ val1 + val2) >>>>>>>> >>>>>>>> I get back an immediate error that tells me, "The number of >>>>>>>> alternatives should be at least 3" >>>>>>>> >>>>>>>> Since I have a binary training label, this looks like a >>>>>>>> problem. (Additionally, I thought that a probit was a >>>>>>>> appropriate tool for building binary models.) >>>>>>>> >>>>>>>> Any advice? >>>>>>> >>>>>>> *Multi*nomial probit typically conveys more than 2 choices while >>>>>>> *bi*nomial probit conveys exactly 2 choices. One could argue >>>>>>> that the latter should be a special case of the former but the >>>>>>> more general case has much more computational challenges. >>>>>>> >>>>>>> The =2 vs >2 information might have been inferred from the title >>>>>>> of the package already but if you wanted to take extreme actions >>>>>>> you could read the mnp() manual page or oven the references it >>>>>>> points you to: The software is discussed in the Journal of >>>>>>> Statistical Software (http://www.jstatsoft.org/v14/i03/) and the >>>>>>> theory is described in an article in the Journal of Econometrics >>>>>>> (124, 311-334). >>>>>>> >>>>>>> Z >>>>>>> >>>>>>>> Thanks! >>>>>>>> >>>>>>>> -N >>>>>>>> >>>>>>>> ______________________________________________ >>>>>>>> R-help@r-project.org mailing list >>>>>>>> https://stat.ethz.ch/mailman/listinfo/r-help >>>>>>>> PLEASE do read the posting guide >>>>>>>> http://www.R-project.org/posting-guide.html >>>>>>>> and provide commented, minimal, self-contained, reproducible code. >>>>>>>> >>>>>>>> >>>>>> >>>> >> [[alternative HTML version deleted]] ______________________________________________ R-help@r-project.org mailing list https://stat.ethz.ch/mailman/listinfo/r-help PLEASE do read the posting guide http://www.R-project.org/posting-guide.html and provide commented, minimal, self-contained, reproducible code.
