Thanks. But please trust me, I'm not making an elementary error. I
know that if you center the variables, the intercept will change from
the non-centered situation. To summarize the problem, when jmp centers
the variables (as it does via default), one gets a different intercept
than when i do the centering myself.
On May 10, 2009, at 8:49 PM, Friedman, Emil wrote:
Let's consider a simpler example. Imagine the two equations:
y = b0 + b1*X versus y = b'0 + b'1(X-Xbar)
where Xbar is the mean of the X's.
The two equations will make identical predictions if
b1 = b'1 and b0 = b'0 - b'1*Xbar
but b0 will not equal b'0 so the two approaches will give us
different intercepts.
It's like going from Fahrenheit to Celcius. The intercept on the
Celcius scale is at freezing point of water. The intercept on the
Fahrenheit scale is at (I've forgotten but I think it was whatever
the coldest achievable temperature was for the guy who invented it.)
Does that help?
----------------------------------------------
Emil M Friedman, PhD
www.statisticalconsulting.org <http://www.statisticalconsulting.org/>
emilfriedman@sbcglobal.net
efriedman@mannkindcorp.com
65 Cables - Unit 6
Waterbury, CT 06710
216-287-0821 (cell)
MannKind Biopharmaceuticals
(Non-Clinical) Statistician
One Casper Street
Danbury, CT 06810
203-790-2507 (office)
________________________________
From: jmp-l-owner@lists.biostat.wustl.edu on behalf of Michael Bailey
Sent: Sun 5/10/2009 5:55 PM
To: jmp-l@lists.biostat.wustl.edu
Subject: Re: [jmp-l] intercept in jmp 7, "fit model?"
Thanks to Al Best for his suggestion. However, I cannot find any
information about the intercept. I understand about centering
continuous predictors in a polynomial context. The regression
coefficients all make sense. What doesn't make clear sense is the
value of the Y intercept.
For example, if I use a sample data set (Facsalarydummyexamp.jmp)
(Download:
http://www.psych.northwestern.edu/psych/people/faculty/bailey/Facsalarydummyexamp.jmp)
if I regress Salary on Merit, Years, and their interaction, using
the cross button in the "fit model" box, I get 55.305 for the
intercept. If, however, I do this by first creating deviation scores
by hand and then creating a product of deviation variables for the
interaction term, I get exactly the same value for the regression
coefficients--except for the intercept, which is now 50.499. I have no
idea where that number comes from.
On May 10, 2009, at 3:19 PM, Al Best wrote:
center polynomials
Michael Bailey
jm-bailey@northwestern.edu
<winmail.dat>
Michael Bailey
jm-bailey@northwestern.edu
|