Re: Enter versus forward method for linear regression

Dear Richard.

Thanks for your help. If you don't mind I will tell you a bit more
about my dataset and see what you think.

I have a sample of about 300, although for much of the analysis this is
split into 2 groups analysed independently. I prob have about 10
predictor variables for many of the dependents, which as far as I know
does not violate basic guides for the number permitted for the sample
size. All of these are based on evidence of associations or at least
related to the dependent theoretically.

I am using SPSS, but am writing up away from college so do not have
that much access to text books. I am also on a tight schedule so need
to proceed with this analysis asap

From what you said I plan to continue as follows.

Use the enter method for all blocks
Start with my primary predictor variable in block 1.
Sequentially add known confounders in blocks, e.g age and sex in block
2.
Examine what happens to the relationship between the dependent and
primary predictor when additional predictor varibles are added. i.e
what happens to the size and direction of the coefficients and the
significance level.

One of my concerns with proceeding in this manner is that as I said I
will have many predictors whose relationships with the dependent are
not significant. I would still like to present these as I think it is
important to do so, however as I understand it my R squared value will
not be truly reflective of the fianl model because of the extra
predictors. This is why I went on to do the forward (not stepwise)
method. However this alone tells me nothing about my primary predictor
of interest as it is often not significantly related. The outcome is
the same with the enter method, but my feeling is it provides more
information when looking at a specific predictor rather than trying to
determine the best model.

Lastly I will repeat the regression with and without the variables that
I think are involved in the mechanism and see what that does to the
relationships.

Thanks again
J

.

Relevant Pages
Quantcast