|
You can also do the analysis in the presence of the
singularities. **If** the interaction term turns
out to be insignificant, you could remove it and re-run the model. The
singularity will disappear when you remove the interaction term.
----- Original Message -----
Sent: Thursday, October 20, 2005 9:22
AM
Subject: Re: [jmp-l] Is Nesting
Appropriate?
Sean,
I don't think nesting is appropriate because the
levels of B do not vary by A. In other words "+" and "-" mean the same
settings regardless of A's value.
I worked with your data a bit and
suggest the following:
1. Start with YbyX plots of Y vs A and Y vs B
(hint; color code rows for B ).
2. Modeling: I would think it would
be best to leave 5 and 6 out to model A, B, and A*B. Use Effect
Screening and checkout the profile plots.
3. Test A levels 5 and 6
with the other levels at only B = "-".
4. If you were to find that
there is no A*B interaction over A levels 1-4 and that B has no effect,
then you could consider using all data for the comparison of A levels since
the "+" B's would simply be additional replicates at the A
levels.
Michael Bresnick
BMS Medical Imaging
Davern, Sean
wrote:
I've completed the following screening experiment:
Factor A (categorical): 6 levels
Factor B (categorical): 2 levels: + / -
Two levels of factor A (5 & 6) are internal controls linking this
experiment's results to other work. However, those levels of factor A were
only run at the - level of factor B since that's how they were run in the
other work. Each treatment was run twice (1 replicate). So, here's the
(un-randomized) design (without the replicates):
A B
1 +
2 +
3 +
4 +
1 -
2 -
3 -
4 -
5 -
6 -
If I analyze the data with
Fit Model(Effects( :A, :B, :A* :B), Y(:Y))
there will be errors (singularities) because the 5+ and 6+ treatments don't
exist. So, I can analyze the data excluding all runs of level 5 and 6.
However, then if I want to know if the performance of runs 5+ and 6+ are
statistically distinguishable from any or all the others (using Tukey's HSD)
I don't know how to get the result.
I'm wondering if I can use the nested model:
Fit Model(Y( :Y), Effects( :A, :B[ :A]), Personality(Standard Least
Squares))
to do the analysis. I believe it is calculating the means correctly but I'm
concerned that it will have incorrect estimates of variances and degrees of
freedom and that the hypothesis testing will be wrong. It certainly gives
different estimates of MSE, etc. but I'm not sure if that is because it now
has more data to work with.
Is use of the nested model appropriate to compare the performance at each
treatment as well as the parameter estimations?
Here is the JMP file including models and scripts for the fits:
<<Experiment Design and Data.JMP>>
Thanks, in advance, for your help.
Sean
Sean Davern
Engineer III
Cell Sciences Process Development
Mail Stop AW2/D2152
Ext. 57074
|