[EMAIL PROTECTED] wrote:
> My response variable is continuous and ranges betwen 0.77 to 1.2.
> I have 31 observations, 6 factors each at 3 levels. Its not a full
> factorial, since that would require 729 runs. My full model contains
> all the main effects (which would be 6 terms), the pair wise second
> order interaction terms(which would be 15 terms) and the square terms ,
> i.e. x1*x1 x2*X2 X3*X3 X4*X4 X5*X5 and X6*X6.
> When I did the stepwise regression I was able to get rid of X1*X2 ,
> X1*X3, X2*X5 , X4*X4 and X5*X5(I apologize the error in my post, which
> says I was able to get rid of only 3 variables) So, I basically have 8
> df for my error term.
Isn't there a paper by Box along the lines of
- putting non-significant effects into the error term will cause the
standard error to be under-estimated
- leading to more terms being kept in the model then should be there.
IIRC the recommendation was to use a half-normal plot of the effect
estimates.
However if all those estimates are significant then a half normal plot
wont help. Have a look and see anyway.
Is it possible to do a replicate?
M.
=================================================================
Instructions for joining and leaving this list and remarks about
the problem of INAPPROPRIATE MESSAGES are available at
http://jse.stat.ncsu.edu/
=================================================================
