I do appreciate this answer. I heard that in SAS, conditional logistic model do predictions in the same way. However, this formula can only deal with in-sample predictions. How about the out-of-sample one? Is it like one of the former responses by Thomas, say, it's impossible to do the out-of-sample prediction??
Results from clogit out of range?
5 messages · David Winsemius, Lisa, Jay
I still don't think the exp(lp)/(1+exp(lp)) gonna work. Since this is conditional logit model, while this formula is only used in unconditional ones. By using this, one neglects the information based on stratum. Though I don't know how to solve it to. I am also working on a project on this and I do hope there's someone explaining this problem. Will that be a possibility that the phat can never be estimated as we never know the individual intercept?
This appears to be addressed to a thread that appeared almost three years ago. I suspect you have not read all the way to the end of the thread: https://stat.ethz.ch/pipermail/r-help/2010-April/235956.html
On Feb 28, 2013, at 5:45 PM, lisa wrote:
I do appreciate this answer. I heard that in SAS, conditional logistic model do predictions in the same way. However, this formula can only deal with in-sample predictions. How about the out-of-sample one? Is it like one of the former responses by Thomas, say, it's impossible to do the out-of-sample prediction??
I do not understand how an "out-of-sample" conditional estimate makes any logical sense. The questioner was asking why he was getting values outside the range of [0,1] which is not the same as asking for estimates for strata outside the range of the stratum values. Charles Berry gave another sensible answer, but I did not interpret his suggestion as solving what you seem to be requesting.
David Winsemius Alameda, CA, USA
Sorry I was asking a question inspired by the original question, which relates to the out of sample prediction. He was asking why it's not of the probability range, and I also found that clogit only gives out linear predictors in both estimation and prediction, but never the probability. For in sample ones, we can derive it as the last thread says; that's what inspired me of thinking of out of sample cases. And I am guessing but not sure if the probability of that case can also be deduced. I'd like to hear of your opinions. Thanks anyway for your response and sorry for the misunderstandings. Thanks, Lisa
On 1 Mar, 2013, at 10:10 PM, David Winsemius <dwinsemius at comcast.net> wrote:
I still don't think the exp(lp)/(1+exp(lp)) gonna work. Since this is conditional logit model, while this formula is only used in unconditional ones. By using this, one neglects the information based on stratum. Though I don't know how to solve it to. I am also working on a project on this and I do hope there's someone explaining this problem. Will that be a possibility that the phat can never be estimated as we never know the individual intercept?
This appears to be addressed to a thread that appeared almost three years ago. I suspect you have not read all the way to the end of the thread: https://stat.ethz.ch/pipermail/r-help/2010-April/235956.html
On Feb 28, 2013, at 5:45 PM, lisa wrote:
I do appreciate this answer. I heard that in SAS, conditional logistic model do predictions in the same way. However, this formula can only deal with in-sample predictions. How about the out-of-sample one? Is it like one of the former responses by Thomas, say, it's impossible to do the out-of-sample prediction??
I do not understand how an "out-of-sample" conditional estimate makes any logical sense. The questioner was asking why he was getting values outside the range of [0,1] which is not the same as asking for estimates for strata outside the range of the stratum values. Charles Berry gave another sensible answer, but I did not interpret his suggestion as solving what you seem to be requesting. -- David Winsemius Alameda, CA, USA
On Mar 1, 2013, at 6:22 AM, Lisa Sheng wrote:
Sorry I was asking a question inspired by the original question, which relates to the out of sample prediction. He was asking why it's not of the probability range, and I also found that clogit only gives out linear predictors in both estimation and prediction, but never the probability.
I would not have expected a function named 'clogit' to produce any of that. I would have guessed that `clogit` from whatever package is under consideration to be either a regression function or a transformation function. It is the 'predict' methods in R that might (or might not) have a "response" or "risk" type of argument to control their output. You are not providing any code so the nature of your concerns seems remain impossibly vague. If your concern were for the output of predict( survival::clogit(form, dataset), type="risk") then you do need to understand that output from survival models are generally relative risks and not probability estimates.
David. > For in sample ones, we can derive it as the last thread says; that's what inspired me of thinking of out of sample cases. And I am guessing but not sure if the probability of that case can also be deduced. > > I'd like to hear of your opinions. Thanks anyway for your response and sorry for the misunderstandings. > > Thanks, > Lisa > > On 1 Mar, 2013, at 10:10 PM, David Winsemius <dwinsemius at comcast.net> wrote: > >>>>> I still don't think the exp(lp)/(1+exp(lp)) gonna work. Since this is conditional logit model, while this formula is only used in unconditional ones. By using this, one neglects the information based on stratum. Though I don't know how to solve it to. I am also working on a project on this and I do hope there's someone explaining this problem. Will that be a possibility that the phat can never be estimated as we never know the individual intercept? >> >>>> This appears to be addressed to a thread that appeared almost three years ago. I suspect you have not read all the way to the end of the thread: >>>> >>>> https://stat.ethz.ch/pipermail/r-help/2010-April/235956.html >> On Feb 28, 2013, at 5:45 PM, lisa wrote: >> >>> I do appreciate this answer. I heard that in SAS, conditional logistic model do predictions in the same way. However, this formula can only deal with in-sample predictions. How about the out-of-sample one? Is it like one of the former responses by Thomas, say, it's impossible to do the out-of-sample prediction?? >> >> I do not understand how an "out-of-sample" conditional estimate makes any logical sense. The questioner was asking why he was getting values outside the range of [0,1] which is not the same as asking for estimates for strata outside the range of the stratum values. Charles Berry gave another sensible answer, but I did not interpret his suggestion as solving what you seem to be requesting. >> >> -- >> >> David Winsemius >> Alameda, CA, USA >> David Winsemius Alameda, CA, USA
I'm not positive of the question you are asking b/c I lost some of initial messages in thread but I think
predict(model, type="expected")
gives fitted probabilities Apologies if I answered a question no one asked.
On Feb 28, 2013, at 7:45 PM, lisa <lisha.sheng at gmail.com> wrote:
I do appreciate this answer. I heard that in SAS, conditional logistic model do predictions in the same way. However, this formula can only deal with in-sample predictions. How about the out-of-sample one? Is it like one of the former responses by Thomas, say, it's impossible to do the out-of-sample prediction??
______________________________________________ R-help at 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.