R help - May 2009 - Error with regsubset in leaps package - vcov and all.best option (plus calculating VIFs for subsets)

Hi all


I am hoping this is just a minor problem, I am trying to implement a best
subsets regression procedure on some ecological datasets using the regsubsets
function in the leaps package. The dataset contains 43 predictor variables plus
the response (logcount) all in a dataframe called environment. I am implementing
it as follows:

library(leaps)


subsets<-regsubsets(logcount~.,data=environment,nvmax=10,nbest=2,really.big=FALSE,method="exhaustive")

###the subset regression runs fine when i run it as above and i can get all the
usual summaries

###The problem comes when i try and get it to output the variance convariance
matric by adding the option vcov=TRUE
##When I do that i get the following:

subsets<-regsubsets(logcount~.,data=environment,nvmax=10,nbest=2,really.big=FALSE,method="exhaustive",vcov=TRUE)

Error in model.frame.default(data = environment, vcov = TRUE, formula =
logcount~  :
  variable lengths differ (found for '(vcov)')
> traceback()
6: model.frame.default(data = environment, vcov = TRUE, formula = logcount~ 
       .)
5: model.frame(data = environment, vcov = TRUE, formula = logcount~ .)
4: eval(expr, envir, enclos)
3: eval(mm, sys.frame(sys.parent()))
2: regsubsets.formula(logcount~ ., data = environment, nvmax = 10, really.big =
TRUE,
       method = "exhaustive", nbest = 5, vcov = TRUE)
1: regsubsets(obs ~ ., data = environment, nvmax = 10, really.big = TRUE, 
       method = "exhaustive", nbest = 5, vcov = TRUE)

I get the same error when i try and add the all.best=TRUE option (as ideally i
would like it to report the fits of all the subsets).

All the predictor variables and the response are the same length (143) so not
sure if I am misinterpreting the error or have misspecified the regsubsets?

I was wanting to get the variance - covariance matrix as I believe I need it to
calculate the Variance Inflation Factors (VIFs) for each of the models reported
by regsubsets. As I want to exclude any models that exhibit multicollinerarity
from later analysis, I am hoping to select say the 'best' 10 models and
bootstrap them to find out more about how they perform.

Or am I going about this all the wrong way? is there away to calulate vifs from
regsubsets or pass it directly to something that calculates them e.g. VIF in the
car package?

Any help will be most appreciated, many thanks in advance

Lara

lara.harrup at bbsrc.ac.uk

On Wed, 20 May 2009, lara harrup (IAH-P) wrote:
>
>
> Hi all
>
>
> I am hoping this is just a minor problem, I am trying to implement a best
subsets regression procedure on some ecological datasets using the regsubsets
function in the leaps package. The dataset contains 43 predictor variables plus
the response (logcount) all in a dataframe called environment. I am implementing
it as follows:
>
> library(leaps)
>
>
>
subsets<-regsubsets(logcount~.,data=environment,nvmax=10,nbest=2,really.big=FALSE,method="exhaustive")
>
> ###the subset regression runs fine when i run it as above and i can get all
>the usual summaries
>
> ###The problem comes when i try and get it to output the variance
convariance >matric by adding the option vcov=TRUE
Yes, that would be because there is no option vcov=TRUE for regsubsets. There is
a vcov= option for the coef() method, which may be what is confusing you.

<snip>
> I was wanting to get the variance - covariance matrix as I believe I need
it >to calculate the Variance Inflation Factors (VIFs) for each of the models
>reported by regsubsets. As I want to exclude any models that exhibit
>multicollinerarity from later analysis, I am hoping to select say the
'best' >10 models and bootstrap them to find out more about how they
perform.
As in the example on the help page, once you have run regsubsets() you can use 
coef() and vcov() on the object it returns to get coefficient estimates and
variance-covariance matrices for any of the best models.

      -thomas

Thomas Lumley			Assoc. Professor, Biostatistics
tlumley at u.washington.edu	University of Washington, Seattle