R help - Apr 2004 - Non-homogeneity of variance - decreasing variance

Hello all,
I'm running very simple regression but face a problem of non-homogeneity of
variance, but with a decreasing variance with increasing mean...I do not
know how to deal with that.
this relationship doesn't seem to be strong, but it's my first time to
see
something like that, and would like to know what to do if one day it becomes
stronger. I tested just for fun some transformation but was not able to get
a better model. I do not know if it can help, but my predictor variable is a
kind of gamma poisson-shaped-like zero-rich distribution (continuous of
course), highly overdispersed.
If one know how to deal with decreasing variance, I would appreciate any
advice (I tried to modelize negative variance-mean relationship in a new
quasi- family this was prohibited, only constant, mu, mu^x (and mu(1-mu) for
binomial) were allowed). I've definitively reached the border of the
statistical black box for me.
thanks
simon

	[[alternative HTML version deleted]]

On 13 Apr 2004 at 19:36, Simon Chamaill?? wrote:

You could maybe try gls in package nlme, where you can estimate
 parameters in variance functions. If you need a generalized linear 
model, you could have a look at glmmPQL in MASS, but I don't know if 
that accepts models without random effects.

Kjetil Halvorsen
> Hello all,
> I'm running very simple regression but face a problem of
> non-homogeneity of variance, but with a decreasing variance with
> increasing mean...I do not know how to deal with that. this
> relationship doesn't seem to be strong, but it's my first time to
see
> something like that, and would like to know what to do if one day it
> becomes stronger. I tested just for fun some transformation but was
> not able to get a better model. I do not know if it can help, but my
> predictor variable is a kind of gamma poisson-shaped-like zero-rich
> distribution (continuous of course), highly overdispersed. If one know
> how to deal with decreasing variance, I would appreciate any advice (I
> tried to modelize negative variance-mean relationship in a new quasi-
> family this was prohibited, only constant, mu, mu^x (and mu(1-mu) for
> binomial) were allowed). I've definitively reached the border of the
> statistical black box for me. thanks simon
> 
>  [[alternative HTML version deleted]]
> 
> ______________________________________________
> R-help at stat.math.ethz.ch mailing list
> https://www.stat.math.ethz.ch/mailman/listinfo/r-help
> PLEASE do read the posting guide!
> http://www.R-project.org/posting-guide.html

Dear Simon,

I'm not sure that I follow this entirely, but if error variance decreases
with the level of the response, you could try raising the response to a
power greater than 1. Of course, the response has to be non-negative. You
might take a look at the spread.level.plot function in the car package,
which will produce a suggested transformation when applied to an lm object.

I hope that this helps,
 John 
> -----Original Message-----
> From: r-help-bounces at stat.math.ethz.ch 
> [mailto:r-help-bounces at stat.math.ethz.ch] On Behalf Of Simon Chamaill??
> Sent: Tuesday, April 13, 2004 12:36 PM
> To: r-help at stat.math.ethz.ch
> Subject: [R] Non-homogeneity of variance - decreasing variance
> 
> Hello all,
> I'm running very simple regression but face a problem of 
> non-homogeneity of variance, but with a decreasing variance 
> with increasing mean...I do not know how to deal with that.
> this relationship doesn't seem to be strong, but it's my 
> first time to see something like that, and would like to know 
> what to do if one day it becomes stronger. I tested just for 
> fun some transformation but was not able to get a better 
> model. I do not know if it can help, but my predictor 
> variable is a kind of gamma poisson-shaped-like zero-rich 
> distribution (continuous of course), highly overdispersed.
> If one know how to deal with decreasing variance, I would 
> appreciate any advice (I tried to modelize negative 
> variance-mean relationship in a new
> quasi- family this was prohibited, only constant, mu, mu^x 
> (and mu(1-mu) for
> binomial) were allowed). I've definitively reached the border 
> of the statistical black box for me.
> thanks
> simon
>