From: Frank E Harrell Jr <f.harrell_at_vanderbilt.edu>

Date: Thu, 29 May 2008 07:44:47 -0500

>> Xiaohui Chen wrote:

*>>> step or stepAIC functions do the job. You can opt to use BIC by
*

*>>> changing the mulplication of penalty.
*

*>>>
*

*>>> I think AIC and BIC are not only limited to compare two pre-defined
*

*>>> models, they can be used as model search criteria. You could
*

*>>> enumerate the information criteria for all possible models if the
*

*>>> size of full model is relatively small. But this is not generally
*

*>>> scaled to practical high-dimensional applications. Hence, it is often
*

*>>> only possible to find a 'best' model of a local optimum, e.g.
*

*>>> measured by AIC/BIC.
*

*>>
*

*>> Sure you can use them that way, and they may perform better than other
*

*>> measures, but the resulting model will be highly biased (regression
*

*>> coefficients biased away from zero). AIC and BIC were not designed to
*

*>> be used in this fashion originally. Optimizing AIC or BIC will not
*

*>> produce well-calibrated models as does penalizing a large model.
*

*>>
*

Date: Thu, 29 May 2008 07:44:47 -0500

Xiaohui Chen wrote:

> Frank E Harrell Jr 写道:

>> Xiaohui Chen wrote:

> Sure, I agree with this point. AIC is used to correct the bias from the > estimations which minimize the KL distance of true model, provided the > assumed model family contains the true model. BIC is designed for > approximating the model marginal likelihood. Those are all > post-selection estimating methods. For simutaneous variable selection > and estimation, there are better penalizations like L1 penalty, which is > much better than AIC/BIC in terms of consistency.

The main point is to use some kind of penalization. Lots of people are using AIC/BIC to select models using ordinary unpenalized least squares or maximum likelihood, and that doesn't work well.

*>>>
*

>>> On the other way around, I wouldn't like to say the over-penalization

*>>> of BIC. Instead, I think AIC is usually underpenalizing larger models
**>>> in terms of the positive probability of incoperating irrevalent
**>>> variables in linear models.
**>>
**>> If you put some constraints on the process (e.g., if using AIC to find
**>> the optimum penalty in penalized maximum likelihood estimation), AIC
**>> works very well and BIC results if far too much shrinkage
**>> (underfitting). If using a dangerous process such as stepwise
**>> variable selection, the more conservative BIC may be better in some
**>> sense, worse in others. The main problem with stepwise variable
**>> selection is the use of significance levels for entry below 1.0 and
**>> especially below 0.1.
*

> What's the point to use AIC on penalized MLE? I think generally you can > view the penalty as the prior regularization and using certain > optimization algorithm to find the MAP estimate.

Simulations show that AIC can choose the optimum L2 penalty in PMLE using the effective degrees of freedom of each model. Optimizing BIC choose far too much shrinkage.

Frank

*>>
**>> Frank
**>>
**>>>
**>>> X
**>>>
*

>>> Frank E Harrell Jr 写道:

*>>>> Smita Pakhale wrote:
**>>>>> Hi Maria,
**>>>>>
**>>>>> But why do you want to use forwards or backwards
**>>>>> methods? These all are 'backward' methods of modeling.
**>>>>> Try using AIC or BIC. BIC is much better than AIC.
**>>>>> And, you do not have to believe me or any one else on
**>>>>> this.
**>>>>
**>>>> How does that help? BIC gives too much penalization in certain
**>>>> contexts; both AIC and BIC were designed to compare two
**>>>> pre-specified models. They were not designed to fix problems of
**>>>> stepwise variable selection.
**>>>>
**>>>> Frank
**>>>>
**>>>>>
**>>>>> Just make a small data set with a few variables with
**>>>>> known relationship amongst them. With this simulated
**>>>>> data set, use all your modeling methods: backwards,
**>>>>> forwards, AIC, BIC etc and then see which one gives
**>>>>> you a answer closest to the truth. The beauty of using
**>>>>> a simulated dataset is that, you 'know' the truth, as
**>>>>> you are the 'creater' of it!
**>>>>>
**>>>>> smita
**>>>>>
**>>>>> --- Charilaos Skiadas <cskiadas_at_gmail.com> wrote:
**>>>>>
**>>>>>> A google search for "logistic regression with
**>>>>>> stepwise forward in r" returns the following post:
**>>>>>>
**>>>>>>
**>>>>> https://stat.ethz.ch/pipermail/r-help/2003-December/043645.html
**>>>>>> Haris Skiadas
**>>>>>> Department of Mathematics and Computer Science
**>>>>>> Hanover College
**>>>>>>
**>>>>>> On May 28, 2008, at 7:01 AM, Maria wrote:
**>>>>>>
**>>>>>>> Hello,
**>>>>>>> I am just about to install R and was wondering
**>>>>>> about a few things.
**>>>>>>> I have only worked in Matlab because I wanted to
**>>>>>> do a logistic
**>>>>>>> regression. However Matlab does not do logistic
**>>>>>> regression with
**>>>>>>> stepwiseforward method. Therefore I thought about
**>>>>>> testing R. So my
**>>>>>>> question is
**>>>>>>> can I do logistic regression with stepwise forward
**>>>>>> in R?
**>>>>>>> Thanks /M
**>>>>>> ______________________________________________
**>>>>>
**>>>>
**>>>
**>>>
**>>
**>>
*

>

>

-- Frank E Harrell Jr Professor and Chair School of Medicine Department of Biostatistics Vanderbilt University ______________________________________________ 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.Received on Thu 29 May 2008 - 12:59:23 GMT

Archive maintained by Robert King, hosted by
the discipline of
statistics at the
University of Newcastle,
Australia.

Archive generated by hypermail 2.2.0, at Thu 29 May 2008 - 13:30:41 GMT.

*
Mailing list information is available at https://stat.ethz.ch/mailman/listinfo/r-help.
Please read the posting
guide before posting to the list.
*