+ Reply to Thread
Results 1 to 5 of 5

Thread: Hypothesis for logistic regression

  1. #1
    Points: 3,866, Level: 39
    Level completed: 44%, Points required for next Level: 84

    Posts
    28
    Thanks
    5
    Thanked 0 Times in 0 Posts

    Hypothesis for logistic regression




    I get a group of data and would like to use the regression to analysis the relationship. The dependent variable Y is a kind of loss rate, so it is strictly belong to [0, 1], e.g. I have $100, and I will loss 5% at day 1, 3% on day 2, 4% day 3, ...then Y is (5%, 3%, 4%, ....); I also have some independent variables X1, X2, etc.

    Cause Y is belong to [0, 1], when I try the linear regression, I get a lot of negative predicted value when doing the forecast, so I would like to try the logistic regression.

    The problem I meet is that, is there any hypothesis for using logistic regression? I choose the family as binomial, the model result is what I would like to see, but I remember that binomial is for 0, 1(classified) problem, can I model the loss rate this way here? Do I need to do any hypothesis test?

    Thanks for the kindly help.

  2. #2
    Omega Contributor
    Points: 38,396, Level: 100
    Level completed: 0%, Points required for next Level: 0
    hlsmith's Avatar
    Location
    Not Ames, IA
    Posts
    7,001
    Thanks
    398
    Thanked 1,186 Times in 1,147 Posts

    Re: Hypothesis for logistic regression

    Rates can be modelled with Poisson. Though your DV is still unclear to me. Is it just those 3 values or anything between 0-1?

    Depending on the values of DV I believe in some instances you can log transform it to get a broader scale to work with.

    Side note as you noted, logistic is usually used with binary or multinomial data, which it seemed you don't have.
    Stop cowardice, ban guns!

  3. #3
    Points: 3,866, Level: 39
    Level completed: 44%, Points required for next Level: 84

    Posts
    28
    Thanks
    5
    Thanked 0 Times in 0 Posts

    Re: Hypothesis for logistic regression

    Hi hlsmith,

    Not only 3 DV, I have about 50 values, by using Poisson as the family, do I need to do any hypothesis test for this? e.g. test the distribution of DV to check if it is possion or not.

    For the log transform, I try the transform Z=log(1-Y) and then use the linear regression Z~X1+X2 ..., the problem is on the forecast, I still get negative forecasted loss rate, it is because the independent variable used for the forecast is under stressed.

    e.g. I model loss rate ~ GDP + unemployment; for the forecast we use forecasted GDP and umemployment with very strong stressed (with a very large jump), this result in a negative projected loss rate, and it doesn't make sense, so I am thinking to use logistic regression to make sure the forecast loss rate is between 0-1.


    Quote Originally Posted by pepsico007 View Post
    I get a group of data and would like to use the regression to analysis the relationship. The dependent variable Y is a kind of loss rate, so it is strictly belong to [0, 1], e.g. I have $100, and I will loss 5% at day 1, 3% on day 2, 4% day 3, ...then Y is (5%, 3%, 4%, ....); I also have some independent variables X1, X2, etc.

    Cause Y is belong to [0, 1], when I try the linear regression, I get a lot of negative predicted value when doing the forecast, so I would like to try the logistic regression.

    The problem I meet is that, is there any hypothesis for using logistic regression? I choose the family as binomial, the model result is what I would like to see, but I remember that binomial is for 0, 1(classified) problem, can I model the loss rate this way here? Do I need to do any hypothesis test?

    Thanks for the kindly help.

  4. #4
    Points: 4,664, Level: 43
    Level completed: 57%, Points required for next Level: 86
    kiton's Avatar
    Location
    Corn field
    Posts
    234
    Thanks
    47
    Thanked 51 Times in 46 Posts

    Re: Hypothesis for logistic regression

    You do not seem to have a count model here with a transformed DV. As such, neither Poisson nor NegBin models are appropriate. Instead, if your DV is bounded between 0 and 1 check (A) beta regression, and (B) GLM fractional logit -- these would be appropriate for rates and proportions bounded within 0-1. Note though, one model is inclusive of min and max values, whereas the other one is not.

  5. #5
    Omega Contributor
    Points: 38,396, Level: 100
    Level completed: 0%, Points required for next Level: 0
    hlsmith's Avatar
    Location
    Not Ames, IA
    Posts
    7,001
    Thanks
    398
    Thanked 1,186 Times in 1,147 Posts

    Re: Hypothesis for logistic regression


    Well if you have counts and provide a denominator you can use Poisson for rates, AKA percentages. This is what I was referencing.


    http://www.biostat.umn.edu/~dipankar...lecture_13.pdf
    Stop cowardice, ban guns!

+ Reply to Thread

           




Tags for this Thread

Posting Permissions

  • You may not post new threads
  • You may not post replies
  • You may not post attachments
  • You may not edit your posts






Advertise on Talk Stats