ASA Connect

 View Only
  • 1.  GLM residual deviance formula for logistic model binomial link

    Posted 03-26-2017 16:49



    "Need help with deviance calculation of a Logistic Model with binomial data. Model estimates are computed by a new method and I want to compare GLM using R to my results. I have tails of the data that include no successes and all successes for about 12 points. The Deviance function given in books does not work for 0 successes (0log0 =?) nor for m successes in m trails, m>1, (m-m)* log(m-m) =?. The Deviance formula gives a result without these tails  on my method that is similar to R results (again without the tails).  The degrees of freedom from  R  indicate all data is being used. So what does R do with these tail points in its calculation of Deviance? I need the exact formula."

     I have run the GLM model using R on three data sets and have the resulting residual deviances.  I need my estimates evaluated through the deviance function for comparison. The GLM results are not quite as good as my method using my minimization technique. It is not a complicated process.


    Rudy Gideon
    Retired Professor
    University of Montana
    Missoula, MT

     

    Sent from Mail for Windows 10

     



  • 2.  RE: GLM residual deviance formula for logistic model binomial link

    Posted 03-27-2017 14:15
    0log(0) is to be taken as 1, given that it starts as xlog(theta) for x=0 and theta to be determined as an MLE, which may be 0.

    ------------------------------
    Kenneth Burnham
    Colorado State University
    ------------------------------



  • 3.  RE: GLM residual deviance formula for logistic model binomial link

    Posted 03-27-2017 14:20
    oh dang; my mistake. 0log(0)=0, not 1.

    theta^x in the likelihood is 1 if x=0. Then log(1)=0.


    ------------------------------
    Kenneth Burnham
    Colorado State University
    ------------------------------