Dear All:
I am currently reading a journal
regarding the tolerancing of parts for assembled products. That is, to what
specifications should the components be made to ensure that the assembled
product meets the required standard.
The experimenter has a total of 7 specs
for 3 components to look at and runs a 2^(7-1) experiment. A GLM is estimated
from the experimental data.
Now, the whole point is to identify which
inputs are causing the greatest variability in the output. Intuitively, this
suggests an ANOVA by treatment combination. However, rather than do this the
author uses the GLM in conjunction with the Error Propagation Formula ("EPF")
and the variances of the inputs to find the overall variability in Y and each of
the 7 inputs' contributions to it.
My question is, "Is there a particular
reason for using this method rather than ANOVA"?
ANOVA could have been used to find the
overall varaibility in Y (SStot/N-1) and each of the MStreats could have been
used to give an estimate of the contribution to variability. Does anyone know of
a good reason why the author would instead find this information using the EPF,
which is only an approximate relationship?
Additionally, does anyone know of a good
reference which explains the advantages of the EPF and also states what the
second-order EPF is?
Thanks in advance.
Bryan.