|
|  |
[Date Prev][Date Next][Thread Prev][Thread Next][Date index][Thread index]
st: Re: Re: Anova
.
I have an experiment where a component is stressed (nominally to
the same
Sorry for not being explicit in my description. The component is
strecthed
until it reaches a specified stress, the first value, and then held
for a 4
hours and the stress at this time measured. Due to experimental
constarints
it is not always possible to obtainn this first stress "exactly". The
outcome is the difference between these two stresses. My concern
was that
the value of the first stress may have an effect on the difference
ae a
covariate or confounding effect, and I wanted to estimate or
eliminate this.
Paul
level) and the stress measured a constant time afterwards for 5
different
materials. I wanted to investigate the effect of material on the
difference
in stress giving me the simple model:
anova diff system
The problem is that the initial nominal stress is not constant and
may have
en effect on diff.
How can do I formulate this model? I tried anova diff f1, continuous
(f1)
but f1 now enters the model as a both a main factor and a component of
diff.
Sounds like you've got a factor for material with five levels, and
measurements of stress on each material at two time points. Material
you are treating as a fixed effect since you are specifically
interested in each kind of material. Time as a factor with two levels
you have decided to do away with and treat as a derived difference
measure.
You are concerned about some experimental error in determining the
first time point, which is supposed to represent a baseline. One way
of dealing with this is to hope that error is not systematic, and
examine many examples of each material. If you do this, then you will
need a factor for specimen, nested in kind of material.
modeling difference:
anova difference material / specimen|material /
Even when you do this, if you model the difference between times of
measurement, you can still use the first measurement as a covariate
(I think).
modeling difference with covariate:
anova difference baseline material / specimen|material /, cont(baseline)
If you choose not to model the derived difference variable and
modeled both time points directly, the design is a split plot (I think).
modeling times directly:
anova stress material/specimen|material time material*time/, repeated
(time)
Post-hoc comparisons and contrasts using anova in Stata are a little
under developed, however. For example, it is not possible to specify
an alternate error term when using the matrix syntax of the test
command. Online FAQ show convolutions to test such, using the cell
means anova, but for most users, a cell means anova approach with
complicated mixed model designs, as described in the online FAQ, is
inaccessible. There are two commands written by others that can help
a semi-manual computation of comparisons and contrasts with alternate
error terms (do findit anovacontrast or findit sme). Obviously, I
would prefer Stata incorporate improved versions of these commands. I
cannot recommend Stata to colleagues who wish to analyze their data
by ANOVA, because they invariable need to get beyond the omnibus
tests and to determine simple main effects, comparisons at certain
levels of a factor, etc., using pooled error terms or error terms
specific to each comparison or contrast.
As a comparison, JMP makes it very easy to examine posthoc contrasts
and comparisons for designed experiments with random effects. Even
then, however, JMP seems to restrict the user to pooled error terms.
You can also make good use of Stata's mixed model command xtmixed for
this situation. Each of the models above could be done with xtmixed.
-Dave
*
* For searches and help try:
* http://www.stata.com/support/faqs/res/findit.html
* http://www.stata.com/support/statalist/faq
* http://www.ats.ucla.edu/stat/stata/
