mnrfit and glmfit for binary regression problem
in [Matlab]
|
Prev: Canon command and state names
Next: www.fashion-long-4biz.com supply nike sports shoes,nike jordan shoes,nike
From: ming on 25 Jun 2010 08:05 I have a binary outcome regression problem. the dependent variables (outcome) are binary, e.g. negative and positive, there are 10 independent variable (predictors). What I want to get are the beta (coefficients), i.e. fitting the model given data. I got quite different results by using mnrfit and glmfit, which presented the coefficients with reversed sign and values were scaled by a certain factor. e.g. mnrfit gives: 1st variable: 0.5; 2nd variable: -1; 3rd variable: 2... glmfit gives: 1st variable: -1; 2nd variable: 2; 3rd variable: -4... I cannot figure out which function is more reasonable and should be used. Besides, I am not sure whether mnrfit should be used for binary problem. Please help, thanks!
From: Peter Perkins on 25 Jun 2010 11:16 On 6/25/2010 8:05 AM, ming wrote: > I cannot figure out which function is more reasonable and should be > used. Besides, I am not sure whether mnrfit should be used for binary > problem. Please help, thanks! GLMFIT is specifically intended to fit models for a binary response, and what you most likely want to use. MNRFIT can be used for a binary response, and gives equivalent results. The difference you see in the coefs is likely due to the way you've passed in the response vector. But MNRFIT is intended for the more general case of more than two response levels, and so there are a variety of different models you can fit. Please see the reference page in the doc. McCullagh&Nelder, the book cited in the doc, is an excellent reference for the models that MNRFIT can fit (and also a good reference for GLMs). Hope this helps.
From: ming on 25 Jun 2010 15:45
thanks very much, Peter. |