I agree, a statistician would be helpful to your problem. Or at least,
someone who can weed through some of the alternatives and potholes in
the roads to your solution. (Sorry, had to say that.)
Jim Kroger wrote:
> Hi,
>
> We are using a program called Fidap to do multiple regression on
> neuroimaging data. The problem is nobody here or where the program came
> from understands how to select coefficients, ie., whether and how to make
> them orthogonal. I am not very sophisticated statistically, but can muddle
> through a book if it's not just equations. Would anyone recommend an
> understandable book on multiple regression? Thanks very much for your
> help.
>
> I'll explain the problem a bit in case it helps in understanding what I
> need to learn.
>
> We administer problems of type A or B. Each has two main periods: solve
> and rest. We want to find areas in the brain that respond to a particular
> epoch during each type of problem. So I guess the factors (regressors?)
> would look like (though they would be extended over all trials):
I a factor was the duration of a rest or solve period, this is
interesting. But you have the situation of time sequence, which
standard DoE (the direction you are aiming for) has a hard time handling.
> A solve | rest B solve | rest .......
>
> |~~~~~~~~~~~|____________________________________________________
>
>
> ____________|~~~~~~~~~~~|________________________________________
>
>
> ________________________|~~~~~~~~~~~|____________________________
>
>
> ____________________________________|~~~~~~~~~~~|________________
>
>
> So we select coefficients for these regressors' epochs as such:
>
> 1 0 0 0
> 0 1 0 0
> 0 0 1 0
> 0 0 0 1
>
> We do not know whether these four regressors (can we call them vectors?)
> need to be orthogonal. Especially as we get into more complicated analyses
> where the regressors or effects overlap or coincide. Nobody who uses the
> software we're using has used more than two very simple regressors so
> nobody is able to advise us, so we need to learn about multiple regression
> and figure out what to do ourselves.
> Thanks much.
>
> Jim
In US Industry, there exists a 'what software' syndrome, in which
someone selects a certain software package, before they have explored
the technical problem to solve. Then they wonder why statisticians have
less hair than others - it's been pulled out in frustration. This is
often the level at which the decision maker can ask the questions, but
it is not in a suitable sequence to efficiently solve the technical
problem at hand.
Yes, you will want to do studies that involve orthogonal arrays, for
those question that can be formulated to use them. I wonder about your
issue of order/sequence, nonetheless. If order makes a difference, or
is adjusted, then the order is a factor, and the number of factors you
must consider just blossomed, greatly.
traditional DoE's address a (small) portion of the system at hand, in
such a way that they usually avoid feedback _within_ the experimental
conditions. Only you can understand if this is an issue, although an
experienced DoE person (statistician or otherwise) can often help dig
out the phenomena.
If you are willing to handle the real thing, I'd suggest Box & Draper,
1987, Empirical Model Building & Response Surfaces. A bit heavy on the
math, but watch the pictures. And orthogonal designs are all there.
Jay
--
Jay Warner
Principal Scientist
Warner Consulting, Inc.
4444 North Green Bay Road
Racine, WI 53404-1216
USA
Ph: (262) 634-9100
FAX: (262) 681-1133
email: [EMAIL PROTECTED]
web: http://www.a2q.com
The A2Q Method (tm) -- What do you want to improve today?
=================================================================
Instructions for joining and leaving this list and remarks about
the problem of INAPPROPRIATE MESSAGES are available at
http://jse.stat.ncsu.edu/
=================================================================
