There are quite a number of issues to be dealt with here, and e-mail may
not be the most efficient vehicle for the dealing. Is there not a
statistical consulting service of some description at your university?
(Usually there are several, of varying degrees of formality.)
You may find it instructive to read through my white paper on modelling
and interpreting interactions in multiple regression, on the MINITAB web
site. Go to www.minitab.com and look for "White Papers" -- I've
forgotten the details at this writing.
(Of course, you didn't mention interactions, but if you're doing this
analysis for a thesis, you should probably model at least some of the
possible interactions. But my reasons for suggesting the paper are that
some of your explicit questions are dealt with there. (E.g., with 15
predictors, in the initial analysis one was thrown out by the regression
program and none of the regression coefficients were significant,
although a significant amount of variation in the dependent varibles was
explained by the regression.) That paper is focussed on the idea of
orthogonalizing predictors in a hierarchical way; but the hierarchy
need not (and probably should not) be based solely on the sample values
of correlation coefficients. And you need not intend to orthogonalize
your variables to make sense out of the discussions.)
In fact, I am inclined to suspect that you are misusing "hierarchical"
in your post: it sounds rather as though you are merely ordering the
variables according to their perceived importance (operationalized as
strength of corelation) at the outset. (I take it that you used only
zero-order correlations, not any of the available partial correlations,
for this ordering?)
On Fri, 12 Mar 2004 [EMAIL PROTECTED] wrote:
> I really need a hand with a regression analysis for my thesis. I'll
> try as best I can to summarize/explain. I have a DV, time on task. I
> have a number of IVs, demographics of the participant, cognitive
> tests.
>
> Generated a series of bivariate correlations for time on task. Took
> all significant correlates of Time On Task (11 variables) and entered
> them in a hierarchical regression, the ordering based on the strength
> of the correlation.
This screening may have been a mistake. It is entirely possible for
a highly influential variable to have a non-significant zero-order
correlation with the response variable, due to the effects of its
correlation(s) with other predictor(s). (The phenomenon is called
"suppression", and the other predictors that produce this effect are
called "suppressors". Look 'em up.)
> My problem/question is this: models 1, 2, 3, and 7 are significant. 4,
> 5, 6 are not. What the heck do I report? Do I do a new regression
> based on this? If I report model 7 as significant, do I report that
> the variables added in 4 - 6 were not significant?
This question cannot be usefully addressed, in my experience at any
rate, without having the complete output available to inspect.
Answers, at the rather general level of your questions, are simply not
possible. We have no idea what variables are represented in your
several models, nor what you think their theoretical relationships
should, or might, be. Nor even if your sample size is large enough to
deal with recommendations we might be prepared to suggest.
> And while I am at it, in the coefficients table, one of the variables
> in model seven is listed as significant, and in the final model (all
> 11 variables) only two are significant? What the heck do I report
> here? . .
I rather suspect that you do not understand in adequate detail what
"significant" means in the context of a coefficients table, as distinct
from an analysis-of-variance table. The conditions (under which the
conditional probabilities are assessed) vary from model to model, and
this often leads to the kinds of results that you describe.
("Significance" is a statement of conditional probability.)
Hmm. Possibly you _should_ consider orthogonalizing your predictors,
with respect to a suitable hierarchical order...
------------------------------------------------------------
Donald F. Burrill [EMAIL PROTECTED]
56 Sebbins Pond Drive, Bedford, NH 03110 (603) 626-0816
.
.
=================================================================
Instructions for joining and leaving this list, remarks about the
problem of INAPPROPRIATE MESSAGES, and archives are available at:
. http://jse.stat.ncsu.edu/ .
=================================================================
