Allan ... The "statistical significance" of the regression coefficient
depends on the magnitude of the coefficient and the magnitude of its
standard error (and the appropriate degrees of freedom). So a large
coefficient with a large standard error may very well be less
"significant" than a smaller coefficient. The degree to which
collinearity is confounding the estimates is a separate question.
Art
-----Original Message-----
From: Allan Lundy [mailto:
[hidden email]]
Sent: Wednesday, January 28, 2009 8:22 AM
To:
[hidden email]
Subject: smaller beta weights more significant than larger ones
Greetings Listers!
I have a consulting client who is running multiple regressions with
N=220, with 6 main effect predictors and 20 interactions of those
predictors.
First, I was surprised that he was getting beta weights as high as +/-
3.00
-- I don't recall ever seeing betas above 1.00 before. However, I did
find some examples in published literature, so I am OK with that.
However, he is also getting smaller beta weights, such as .30, that are
highly significant, while larger ones, such as .78, are not. This is
true for both ME and interaction terms. Obviously multicolinearity is
very high, but how weird is this effect with the betas? I did find a
posting, Item # 048894, that mentioned betas of the same size where one
was significant and the other was not, but not the situation I describe.
Thanks a lot in advance!
Allan
=====================
To manage your subscription to SPSSX-L, send a message to
[hidden email] (not to SPSSX-L), with no body text except the
command. To leave the list, send the command SIGNOFF SPSSX-L For a list
of commands to manage subscriptions, send the command INFO REFCARD
=====================
To manage your subscription to SPSSX-L, send a message to
[hidden email] (not to SPSSX-L), with no body text except the
command. To leave the list, send the command
SIGNOFF SPSSX-L
For a list of commands to manage subscriptions, send the command
INFO REFCARD
Locked
