[R] re lative importance of predictors

[Ulrike Grömping]

Hi Robert,

relaimpo does work with a (non-robust) linear model, i.e. calculations in
relaimpo are using correlations between variables, and it is not possible to
incorporate things like the Huber psi function. I am not an expert on robust
regression, but I try to answer your question from the relaimpo point of
view.

If you run relaimpo on an rlm object, it will calculate relative importances
based on the linear model with your prior weights (that you optionally
specify with weights in rlm). 
I suppose it would be more appropriate to run relaimpo on your model with
the weights as determined by rlm (stored in the w component of the output
object, if I correctly interpret the help on rlm); you would have to do that
explicitly, at least currently, since relaimpo uses the weights() function
on the linear model object, and weights(rlm) returns the "weights" component
of the output object. If a relative importance calculation within a linear
model with modified weights (from "w=") is a desirable solution for (at
least special cases of) rlm objects, this can be incorporated into a future
version of relaimpo as the default for these cases. I think it won't be
possible to incorporate psi-functions.

While I am at it: When trying out rlm, I was surprised that specifying "w="
without simultaneously using the "weights="-option apparently interprets w=
as weights=, i.e. the weights component of the resulting rlm object contains
the initial w-values. On the other hand, when specifying both the weights
and the w option, the weights component of the resulting rlm object contains
the weights only and ignores the w-Option. This seems to me somewhat
inconsistent.

Regards, Ulrike


robert.ptacnik wrote:
> 
> Hei Group,
> 
> I want to compare the relative importance of predictors in a multiple
> linear regression y~a+bx1+cx2...
> 
> However, bptest indicates heteroskedasticity of my model. I therefore
> perform a robust regression (rlm), in combination with bootstrapping (as
> outlined in J. Fox, Bootstrapping Regression Models).
> 
> Now I want to compare the relative importance of my predictors. Can I rely
> on the output of 'relaimpo' here? (which takes 'lm', not 'rlm' objects).
> 
> Thanks,
> Robert
> 
> 
> 
> 
> ----------------------------------------------------------------------------------------------------------------------
> NIVAs hovedkontor har flyttet til nye lokaler i CIENS - Forskningssenter
> for miljø og samfunn; Gaustadalléen 21, 0349 Oslo. Meld deg på vårt
> nyhetsbrev på www.niva.no
> ______________________________________________
> 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.
> 
> 

-- 
View this message in context: http://www.nabble.com/relative-importance-of-predictors-tf4897731.html#a14028871
Sent from the R help mailing list archive at Nabble.com.