R help - Jun 2009 - How to exclude insignificant intercepts using "step" function

I posted this question way down at teh end of another thread realted to an
error in step, but that was stupid since it really is another matter
altogether. I should have posted it separately, as I have now done.

The code below creates a data.frame comprising three marginally noisy
surfaces. The code below (including a fix courtesy of David Winsemius that
avoids a step function error through use of the "by" function) returns
significant coefficients for regressions based on factor grp. In the case of
grp A the intercept is not significantly different from zero. However I
can't get step to consider dropping the intercept term. In the attached code
the returned function for grp A is y ~ x1+x2. However I think it should
return y ~ -1 + x1 +x2 (or I guess y ~ 0 + x1 + x2). If I include -1 in the
model designation then no intercept is returned for grps B and C. Am I not
using step properly? Or perhaps there is an alternative methodology I could
use. Any help appreciated.

# y =      x1 +  x2          for grp A
# y = 2 + 2x1 + 4x2          for grp B
# y = 3 + 2x1 + 4x2 + 6x1x2  for grp C
ind <- matrix(runif(200), ncol=2, dimnames=list(c(),
c("x1","x2")))
d1 <- data.frame(ind,
y=ind[,"x1"]+ind[,"x2"]+rnorm(100,0,0.05),
grp=rep("A",100))
d2 <- data.frame(ind,
y=2+2*ind[,"x1"]+4*ind[,"x2"]+rnorm(100,0,0.05),grp=rep("B",
100))
d3 <- data.frame(ind,
y=3+2*ind[,"x1"]+4*ind[,"x2"]+6*ind[,"x1"]*ind[,"x2"]+rnorm(100,0,0.05),grp=rep("C",100))
data2 <- rbind(d1,d2,d3)
# Fit each surface by grp
model <- y ~ x1*x2
by(data2, data2$grp, function(x) {step(lm(model, data=x))}) 

-- 
View this message in context:
http://www.nabble.com/How-to-exclude-insignificant-intercepts-using-%22step%22-function-tp24158818p24158818.html
Sent from the R help mailing list archive at Nabble.com.

I think you should explain (to yourself primarily) what it means to  
have a non-significant intercept. If you can justify on a theoretic  
basis the exclusion of an intercept, then you may get more assistance.  
However, if you are just naively questing after some mythical concept  
of "significance", people may be less motivated to solve what most  
would consider to be an "insignificant" question.

-- 
DW
On Jun 22, 2009, at 11:38 PM, Chris Friedl wrote:
>
> I posted this question way down at teh end of another thread realted  
> to an
> error in step, but that was stupid since it really is another matter
> altogether. I should have posted it separately, as I have now done.
>
> The code below creates a data.frame comprising three marginally noisy
> surfaces. The code below (including a fix courtesy of David  
> Winsemius that
> avoids a step function error through use of the "by" function)
returns
> significant coefficients for regressions based on factor grp. In the  
> case of
> grp A the intercept is not significantly different from zero.  
> However I
> can't get step to consider dropping the intercept term. In the  
> attached code
> the returned function for grp A is y ~ x1+x2. However I think it  
> should
> return y ~ -1 + x1 +x2 (or I guess y ~ 0 + x1 + x2). If I include -1  
> in the
> model designation then no intercept is returned for grps B and C. Am  
> I not
> using step properly? Or perhaps there is an alternative methodology  
> I could
> use. Any help appreciated.
>
> # y =      x1 +  x2          for grp A
> # y = 2 + 2x1 + 4x2          for grp B
> # y = 3 + 2x1 + 4x2 + 6x1x2  for grp C
> ind <- matrix(runif(200), ncol=2, dimnames=list(c(),
c("x1","x2")))
> d1 <- data.frame(ind,
y=ind[,"x1"]+ind[,"x2"]+rnorm(100,0,0.05),
> grp=rep("A",100))
> d2 <- data.frame(ind,
>
y=2+2*ind[,"x1"]+4*ind[,"x2"]+rnorm(100,0,0.05),grp=rep("B",
100))
> d3 <- data.frame(ind,
>
y=3+2*ind[,"x1"]+4*ind[,"x2"]+6*ind[,"x1"]*ind[,"x2"]
> +rnorm(100,0,0.05),grp=rep("C",100))
> data2 <- rbind(d1,d2,d3)
> # Fit each surface by grp
> model <- y ~ x1*x2
> by(data2, data2$grp, function(x) {step(lm(model, data=x))})
>
> -- 

David Winsemius, MD
Heritage Laboratories
West Hartford, CT