[R] penalized maximum likelihood estimation and logistf

Ravi Varadhan rvaradhan at jhmi.edu
Wed Mar 10 06:11:19 CET 2010


Hi,

The problem is actually not in the parameter estimation, but in the confidence interval computation using the profile-likelihood method.   This can be easily fixed, but the results may still be not sensible.  So, the best thing to do would be to use the "Wald" confidence interval instead of profile likelihood based CI.  Here is how you can do it:

sd.logr <- logistf(AgeRange ~ I(Age - mean(Age)), pl=FALSE, data=sd)  # this computes Wald C.I.

summary(sd.logr)


Hope this helps,
Ravi.

____________________________________________________________________

Ravi Varadhan, Ph.D.
Assistant Professor,
Division of Geriatric Medicine and Gerontology
School of Medicine
Johns Hopkins University

Ph. (410) 502-2619
email: rvaradhan at jhmi.edu


----- Original Message -----
From: Ravi Varadhan <rvaradhan at jhmi.edu>
Date: Tuesday, March 9, 2010 12:24 pm
Subject: Re: [R] penalized maximum likelihood estimation and logistf
To: xiaoyan yu <xiaoyan.yu at gmail.com>
Cc: r-help at r-project.org


>  Let us look at a piece of the relevant code from "logistf" package:
>  
>          iter <- iter + 1
>          XW2 <- crossprod(x, diag(pi * (1 - pi))^0.5)
>          Fisher <- crossprod(t(XW2))
>          covs <- solve(Fisher)
>          H <- crossprod(XW2, covs) %*% XW2
>          if (firth) 
>              U.star <- crossprod(x, y - pi + diag(H) * (0.5 - 
>                  pi))
>          else U.star <- crossprod(x, y - pi)
>          delta <- as.vector(covs %*% U.star)
>          mx <- max(abs(delta))/maxstep
>          if (mx > 1) 
>              delta <- delta/mx
>   
>  First, note that this is inside an iterative loop.  The Fisher 
> information matrix is first formed, and then it is inverted to get 
> `covs'.  This is where it can get troublesome.  It is very likely in 
> your case that the `Fisher' information matrix is ill-conditioned, and 
> hence `covs' cannot be computed.  Consequently, the steplength `delta' 
> for the Newton method cannot be determined. 
>  
>  
>  Ravi.
>  
>  ____________________________________________________________________
>  
>  Ravi Varadhan, Ph.D.
>  Assistant Professor,
>  Division of Geriatric Medicine and Gerontology
>  School of Medicine
>  Johns Hopkins University
>  
>  Ph. (410) 502-2619
>  email: rvaradhan at jhmi.edu
>  
>  
>  ----- Original Message -----
>  From: xiaoyan yu <xiaoyan.yu at gmail.com>
>  Date: Tuesday, March 9, 2010 11:26 am
>  Subject: [R] penalized maximum likelihood estimation and logistf
>  To: r-help at r-project.org
>  
>  
>  > Hi, I got two questions and would really appreciate any help from here.
>  >  First, is the penalized maximum likelihood estimation(Firth Type Estimation)
>  >  only fit for binary response (0,1 or TRUE, FALSE)? Can it be 
> applied 
>  > to
>  >  multinomial logistic regression?
>  >  If yes, what's the formula for LL and  U(beta_i)? Can someone 
> point 
>  > me to
>  >  the right reference?
>  >  Second, when I used *logistf *on a dataset with binary response, I 
> 
>  > got the
>  >  following error msg:
>  >  "Error in if (mx > 1) delta <- delta/mx :
>  >    missing value where TRUE/FALSE needed"
>  >  
>  >  What does this msg mean? does it mean the dataset includes missing 
> response?
>  >  I  checked my dataset several times and didn't find anything other 
> 
>  > than TRUE
>  >  or FALSE for the response. Please shed light on this.
>  >  
>  >  Thanks,
>  >  Xiaoyan
>  >  
>  >  	[[alternative HTML version deleted]]
>  >  
>  >  ______________________________________________
>  >  R-help at r-project.org mailing list
>  >  
>  >  PLEASE do read the posting guide 
>  >  and provide commented, minimal, self-contained, reproducible code.
>  
>  ______________________________________________
>  R-help at r-project.org mailing list
>  
>  PLEASE do read the posting guide 
>  and provide commented, minimal, self-contained, reproducible code.



More information about the R-help mailing list