Re: Variable Importance

On Thu, 14 May 2009 11:05:33 -0700 (PDT), slutsky_fan
<agroeconomist@xxxxxxxxx> wrote:

How do I determine the importance of variables in a logit model- i.e.
which variables contribute the most to explaining the dependent
varaible- is this based on those with the lowest p-value, or the
highest chi-square-

A chi-square with more than 1 d.f. does "measure" a larger
effect, in its way, but the ones with more than 1 d.f. will include
more by chance. Given two factors with the same p-value
(assuming that your program is testing a factor with several d.f.
and providing one test), the one with the large d.f. *is*
accounting for more... as "measured" by the chi-square.

or I've also heard that I should compute
standardized co-efficient s and the variable with the largest co-
efficient is the most important.

The *largest* one corresponds to the best p-value. What
becomes misleading is that the second-largest might be acting
as a suppressor variable, and all the sizes could be distorted
beyond intelligent uni-variable interpretation -- These are
"partial coefficients" in the same sense that applies in multiple
regression, and you can't always interpret a single one by itself.

Or should I complete a varaible
selection procedure and determine 'importance' based on the order in
which the variables enter the model? I've also read that it all
depends on the delta-p value.

Can anyone offer a suggestion?

First, you keep in mind that everything that you are looking
at is what you have obtained "for this sample"; or, by
inference, "for this population."

Second, you need to look at both the univariate tests
and the tests on coefficients in various logit models. To
understand any discrepancies between these, you also
need to look at the intercorrelation of the predictors.

- If the same variable always looks best, then it is the most
important one if all variables are being measured with
equal precision. Or, it is the most important on, given
the validity of measurements. (In other words -- I am trying
to point out that you should not confuse a pragmatic
measurement, which might be inaccurate, with the
philosophical "variable" that shares that name or label.)


- If different variables look best in univariate testing and
in various multiple-variable models, then you are stuck with
weaker statements.

Stepwise entry tells you which one has the highest univariate
relation; and beyond that, it is pretty random. You can
Google groups <group:sci.stat.* stepwise > for additional
comments on that.

--
Rich Ulrich
.

Loading