One should maintain model hierarchy in most situations. In other words, if
there
is an A*B interaction in the model, keep the A and B main effects in the model,
ignore their p-values, and keep in mind that the "physical meaning" of
coefficients change when you change the definition of the dummy variables.
Stepwise and standard least squares use different sets of dummy variables. If
you have a model like A, B, A*B, they will report the same p-value for A*B but
not necessarily the same p-value for A or B. Mark wrote a column about model
hierarchy in the current issue of JMPer Cable
(http://www.jmp.com/about/newsletters/jmpercable/index.shtml).
A similar issue exists for polynomial models (John Nelder, "The Selection of
Terms in Response-Surface Models - How Strong is the Weak-Heredity Principle?",
American Statistician, vol 52, pp 315-318, 1998). Consider a quadratic, Y = b0
+ b1X + b11X^2. If the minimum (or maximum) is near the Y-axis the b1
coefficient will be small. Now change the scale for X (eg, Celcius to Kelvin).
The graph will slide sideways and b1 may well become large.
Hierarchy can be violated when there are good physical reasons for doing so,
but one needs to thoroughly understand the pitfalls.
-------------------------------------------
Emil M Friedman, PhD
18 Clifton Ave
Waterbury, CT 06710
216-287-0821 (cell)
203-790-2507 (work)
2304 Richmond Road
Beachwood, OH 44122
216-591-1750 (Ohio)
emil.friedman@alum.mit.edu
emilfrie@alumni.princeton.edu
http://www.statisticalconsulting.org
-----Original Message-----
From: jmp-l-owner@lists.biostat.wustl.edu
[mailto:jmp-l-owner@lists.biostat.wustl.edu] On Behalf Of Tom Vanwalleghem
Sent: Thursday, July 17, 2008 7:12 AM
To: jmp-l@lists.biostat.wustl.edu
Subject: Re: [jmp-l] interpretation categorical variables
Thanks Emil & Mark for the replies,
I indeed was using the stepwise approach to find the best model fit. I now
tried the standard least squares approach and the problem is some of my
-previously- significant variables are no longer signicant.
I therefore would prefer to keep using the hierarchical approach that is used
by JMP in the stepwise model building.
Is this approach common practice (read: acceptable) when reporting research
results? Truth is I´ve never seen this approach in research articles before.
Thank you,
Tom
-----Mensaje original-----
De: jmp-l-owner@lists.biostat.wustl.edu
[mailto:jmp-l-owner@lists.biostat.wustl.edu] En nombre de Mark Bailey Enviado
el: woensdag 16 juli 2008 23:35
Para: jmp-l@lists.biostat.wustl.edu
Asunto: Re: [jmp-l] interpretation categorical variables
> I´m afraid this could be a basic question, but I did not find the
> answer to my question in the jmp help files, so I turn my hopes to
> this list.
You must be using the Stepwise personality in the Fit Model platform.
If you look in Chapter 19 of the JMP Statistics & Graphics Guide, you will find
an explanation in the Models with Nominal and Ordinal Terms section.
The best that one can do is understand how JMP creates these groups. The
interpretation depends on the nature of the data and your study.
> My variable is soil class (soil - 7 classes). How do I interpret this
> dummy variable JMP creates: " Soil{4&3&1 - 6&7&5} "
It has combined the levels that are most alike (4,3,1) and (6,7,5) into two
groups which are the most different, as far as the response goes.
You can control this behavior and disable it if you find it counter-
productive. Select Rules > Whole Effects. The clusters remain but they now act
as one effect in the fitting process. If you select Make Model, it is copied as
one effect and no columns are created in the data table.
Mark
|