agentsmart wrote:

[...snipped regression results...]

> 1.
> the conclusions based on linear regression and multiple regression are
> totally different.
>
> for instance: X3 seems to be completely unrelated to Y in univariate
> analysis and is the most important important factor in multivariate
> analysis.
>
> obviously, there must be a mistake in my thinking. please correct me.

This can happen, especially if the predictor variables are correlated with

each other. As a extreme example in your field, I cite some old results
according to which the length of a person's left leg is completely
unrelated to the intensity of headaches. Ditto fort he right leg. However,

if you use _both_ variables, there in almost prefect fit. The explanation
is that it is the _difference_ in the length that accounts for the pain.
Univariate analysis would never show that to you.

> 2.
> should i trust the results from linear regression or multiple
> regression?

Just an aside: You are not using the terminology correctly. What
you called "linear regression" is regression with a single predictor,
or "simple regression". "Multiple regression" uses multiple predictors,
but the model is still _linear_ in the coefficients.

> 3.
> which method (univariate regression, multiple regression) should i use
> when both can be used?

Multiple regression is better. But there are many subtle issues of
model building that require careful treatment.

> 4.
> has the parameter "intercept" any value in interpretation of the
> results of multiple regression ?

That depends. The intercept is the estimated mean response when
all the predictor variables are set to 0. If this is a reasonable
combination
of the predictors---sometimes it is, sometimes it isn't---then you can
interpret the intercept as I just did.



.
.
=================================================================
Instructions for joining and leaving this list, remarks about the
problem of INAPPROPRIATE MESSAGES, and archives are available at:
.                  http://jse.stat.ncsu.edu/                    .
=================================================================

Reply via email to