Standard Error in Logistic Regression
Matthias K
works and how that translates into the significance?
The reason for my quest is that we use SPSS for a logistic regression
with extreme weights. My understanding is that SPSS uses the weighted
sample size for calculating standard errors and subsequently the
significances. I would like to use the effective sample size which I
can calculate manually, but I am not sure how to
1. Calculate the standard error for logistic regressions
2. Then get the Wald test from there. Normally I would use the
coefficient and divide it by the s.e. but that doesn't seem to be the
case.
Thanks
Katsche
Rich Ulrich
<katsche.sc...@gmail.com> wrote:
>Can somebody tell me how the standard error in logistic regression
>works and how that translates into the significance?
In general, a standard error comes from a variance divided by an N,
which is degrees of freedom.
Figure it as proportionate to the square root of N.
>
>The reason for my quest is that we use SPSS for a logistic regression
>with extreme weights. My understanding is that SPSS uses the weighted
>sample size for calculating standard errors and subsequently the
>significances. I would like to use the effective sample size which I
>can calculate manually, but I am not sure how to
"Extreme weights" is a worrisome phrase.
A convention for moderate degrees of weighting is to
make the sum of weights equal to the original N. This
preserves an overall approximation that is decent. I
don't know if the logistic routine accepts fractional weights,
if that is what arises.
In very simple situations, with a few categories, the
"effective sample size", total, is probably computed
as a sum of the averages, where the average cell size
in ANOVA is the reciprocal mean.
But "extreme weights", as I read the phrase, really puts
you into a situation of extreme guesswork; any given estimate
depends on some particular number of cases that it rests on,
as a particular contrast; but the overall weighting uses everything.
- You might consider the test on the estimate in the unweighted
model as a decent alternative.... and translate the t-test, if that
is what you have, by inverting to get the confidence limit, if that
is what you are looking for.
- That sounds reasonable to me, but I have no experience with
generating or reading articles with "extreme weights".
>
>1. Calculate the standard error for logistic regressions
>2. Then get the Wald test from there. Normally I would use the
>coefficient and divide it by the s.e. but that doesn't seem to be the
>case.
--
Rich Ulrich
JKPeck
> On Tue, 27 Apr 2010 08:52:49 -0700 (PDT), Matthias K
>
If these weights are sampling weights, the appropriate procedure is
CSLOGISTIC, which knows how to adjust for complex sampling plans. If
these are replication weights, then ordinary logistic is appropriate,
but do be sure that the weights are normalized to the actual number of
cases or the degrees of freedom will be wrong.
Regards,
Jon Peck