[R] Bivariate - multivariate linear regression

[Michael Friendly]

On 5/17/2013 3:45 PM, Jesse Gervais wrote:
> Hi there,
>
>
>
> I want to do several bivariate linear regressions and, than, do a
> multivariate linear regression including only variables significantly
> associated *(p < 0.15)* with y in bivariate analysis, without having to
> look manually to those p values.
>

Rx: You need some stats help here, before getting R help to do something 
that probably doesn't make statistical sense.
Your individual regressions are marginal ones, ignoring other 
predictors. It doesn't
make sense to filter on the marginal analyses before combining into
a multiple (*not* multivariate-- that's when there are several
Ys) regression.

You didn't provide any context, but it would probably be more
useful to proceed in the opposite direction -- fit the model
with all of tol1-tol5 and consider the type II/III tests for
removal (printed as the t-values for coefficients).
Better yet would be to formulate some substantive hypotheses
to test, e.g., using car::linearHypothesis.  For example,
if tol1-tol5 are somehow ordered, e.g., as they might be
if they were polynomial terms in a single predictor, a series
of step-down tests using that order might be appropriate.

You might also look at ?add1, ?drop1 that are designed to do
the kind of sequential testing you were looking for, but these
methods are deprecated when used as automatic procedures to
give you a 'right answer' without any thought.


-- 
Michael Friendly     Email: friendly AT yorku DOT ca
Professor, Psychology Dept. & Chair, Quantitative Methods
York University      Voice: 416 736-2100 x66249 Fax: 416 736-5814
4700 Keele Street    Web:   http://www.datavis.ca
Toronto, ONT  M3J 1P3 CANADA