Skip to content

locfit smoothing question (package maintainer not reachable)

7 messages · Suresh Krishna, David Winsemius, Liaw, Andy

Suresh Krishna
# 
Dear list members,

I am trying to understand this output from the smoothing package locfit  
(1.5-4, running on R 2.8.1 on Windows Vista 64 bit).

# sample code

x<-1:100

y<-rnorm(100)
fit<-locfit(y~x,family="gaussian") #default parameters are fine
plot(fit,band="global")  #plot seems "reasonable", confidence bands use a  
global estimate of variance

y<-1000*rnorm(100)
fit<-locfit(y~x,family="gaussian")
plot(fit,band="global") #aren't these confidence bands too small ? am i  
using this function wrongly ?

Using band="local" gives results that seem to make "more sense". Could  
someone offer me some guidance ?

Thanks, Suresh

ps. The package maintainer, Catherine Loader, is no longer reachable at  
her Auckland address.
Suresh Krishna
# 
Dear all,

I just realized that using family="qgauss" restores normal-looking  
confidence bands... I read that using family="gaussian" rather than  
family="qgauss" fixes the dispersion parameter at 1, but without knowing  
the theory behind the code, I dont understand why there is such a  
difference between the two. If there is a simple explanation or  
recommendation, I am eager to hear it.

Thanks, Suresh


On Tue, 03 Mar 2009 16:56:43 +0100, Suresh Krishna <madzientist at gmail.com>  
wrote:
David Winsemius
# 
I think you should read (or re-read)  the locfit help page and *also*  
the links from that page to the help pages for locfit.raw and rv. I  
would have thought that since family= is not an argument to locfit per  
se, but rather is documented in locfit.raw that you have yet done so,  
but perhaps not?
Liaw, Andy
# 
From: Suresh Krishna
[...]
For the record, I'm the package maintainer for locfit, and I have not
exactly vanished (yet).  Please see the package description.

That said, it doesn't mean I know all the details about the code.  I
just do enough to keep the package on CRAN.

Best,
Andy
Notice:  This e-mail message, together with any attachme...{{dropped:12}}
Suresh Krishna
#
David Winsemis wrote:

            

      
      
      
    

        
I did read the help pages for locfit.raw, and found:

"Local likelihood family; "gaussian"; "binomial"; "poisson"; "gamma" and  
"geom". Density and rate estimation families are "dens", "rate" and  
"hazard" (hazard rate). If the family is preceded by a 'q' (for example,  
family="qbinomial"), quasi-likelihood variance estimates are used.  
Otherwise, the residual variance (rv) is fixed at 1. The default family is  
"qgauss" if a response y is provided; "density" if no response is  
provided. "

However, since the fake data were generated from a known gaussian  
distribution, I did not imagine that using family=gaussian would lead to  
such wildly different results. This is what I was hoping to understand,  
without having to struggle with Catherine's Loader book in order to  
understand the above paragraph deeply enough that this behavior makes  
sense.

Thanks again, Suresh

  
  


  


      

    

    

      
      

        


  

        

      David Winsemius
    

    

      
      #
    

  


  

      
That is what I thought to be the critical paragraph. The variance is  
assumed to be = 1 when you use family="gaussian" rather than the  
default of family="qgauss". You give it a vector, 1000*rnorm(100),  
that ranges widely and a small (relative) variance is assumed and so  
the confidence intervals are plotted as very narrow. This does not  
seem surprising given the functions documented design. I have the book  
and do not think I even need to pull it off the shelf since the help  
pages appear fully informative in this instance. I get an rv of 1 with  
the "gaussian" option and an rv of nearly 1000 when the default is used.

  
    

      
      
    

  


  


      

    

    

      
      

        


  

        

      Suresh Krishna
    

    

      
      #
    

  


  

      
On Tue, 03 Mar 2009 22:10:42 +0100, David Winsemius

        
<dwinsemius at comcast.net> wrote:

        

            

      
      
      
    

        
Thank you, that is helpful. I guess I am wondering under what circumstance  
would it be appropriate to assume that the data had a variance of 1 and  
use the family=gaussian option. Perhaps this is for normalized data ?

Suresh