[R] Regsubsets()

[Thomas Lumley]

On Wed, 11 May 2005, Smit, R. (Robin) (IenT) wrote:

> Dear List members
>
> I am using the regsubsets function to select a few predictor variables
> using Mallow's Cp:
>
>> sel.proc.regsub.full <- regsubsets(CO2 ~ v + log(v) + v.max + sd.v +
> tad + no.stops.km + av.stop.T + a + sd.a + a.max + d + sd.d + d.max +
> RPA + P + perc.stop.T + perc.a.T + perc.d.T + RPS + RPSS + sd.P.acc +
> P.dec + da.acc.1 + RMSACC + RDI + RPSI + P.acc + cov.v + cov.a + cov.d +
> sd.P + sd.v.run + RCS + T + mass.fin, data = DATASET, weights = count,
> nbest = 10, nvmax = 35, method = "exhaustive")
>
> I do however encounter the following warning message which I do not
> understand:
>
>> Reordering variables and trying again:
> Warning messages:
> 1: 14  linear dependencies found in: leaps.setup(x, y, wt = wt, nbest =
> nbest, nvmax = nvmax, force.in = force.in,
> 2: XHAUST returned error code -999 in: leaps.exhaustive(a, really.big
>

The first message means that your variables are linearly redundant (very 
much so, with 14 linear dependencies in 35 variables).  This probably 
explains the second warning.

You need to reduce nvmax: you can't fit any model with more than 35-14=21 
variables. The code will work with linear dependencies as long as you set 
nvmax low enough -- according to Alan Miller's book it was written for 
situtations with more variables than observations.  Try something like 
nvmax=10, and you will be able to see if there really is a use for models 
with even as many as 10 variables.

 	-thomas