F****n
发帖数: 3271 | 1 I have a dependent variable with 2+ categories and I know how to run
multinomial regression with it. However, in multinomial regression you need
to pick a reference category and let other categories compare to it. You don
't get a pairwise comparison. An intuitive solution is to alter the
reference category and run the regression multiple times. However, I am not
quite sure about the validity of this approach. For example, should I worry
about family-wise error? How can I address that? And is there any other
approach available?
Thanks! | y*****n
发帖数: 5016 | 2 how about using the cumulative logit model approach? this approach should
fit your need. | F****n
发帖数: 3271 | 3 Thanks for the reply! But doesn't this require ordinal dependent variables?
My dependent variable is not ordinal.
【在 y*****n 的大作中提到】
: how about using the cumulative logit model approach? this approach should
: fit your need.
| y*****n
发帖数: 5016 | 4 Can you transform the categorical dependent variable into ordinal variable?
I think as long as you can rank order on the categories in terms of
performance, you should be able to make this transformation.
【在 F****n 的大作中提到】
: Thanks for the reply! But doesn't this require ordinal dependent variables?
: My dependent variable is not ordinal.
| F****n
发帖数: 3271 | 5 I hope I could, that would definitely make things much easier. But I don't
think I can do that in this case, as it would be completely meaningless.
Moreover, even you can do so, you get a proportional odds model, which
assume the increase of cumulative probability are uniform across different
ranks.
In multinomial regression, on the contrary, you have one set of
coefficients for each category compared to the reference. And that's what I
want.
?
【在 y*****n 的大作中提到】
: Can you transform the categorical dependent variable into ordinal variable?
: I think as long as you can rank order on the categories in terms of
: performance, you should be able to make this transformation.
| l*********s
发帖数: 5409 | 6 Shall be fine. Plz proceed and let me know the result.
need
don
not
worry
【在 F****n 的大作中提到】
: I have a dependent variable with 2+ categories and I know how to run
: multinomial regression with it. However, in multinomial regression you need
: to pick a reference category and let other categories compare to it. You don
: 't get a pairwise comparison. An intuitive solution is to alter the
: reference category and run the regression multiple times. However, I am not
: quite sure about the validity of this approach. For example, should I worry
: about family-wise error? How can I address that? And is there any other
: approach available?
: Thanks!
| y*****n
发帖数: 5016 | 7 I think you were talking about the score test for the “proportional odds
assumption” , although I am not sure if “the increase of cumulative
probability are uniform across different ranks” is the right way to
describe it. Yes, very often this test will produce p-value below .05, but
in practice, many modelers just don’t care for a variety of reasons.
I
【在 F****n 的大作中提到】
: I hope I could, that would definitely make things much easier. But I don't
: think I can do that in this case, as it would be completely meaningless.
: Moreover, even you can do so, you get a proportional odds model, which
: assume the increase of cumulative probability are uniform across different
: ranks.
: In multinomial regression, on the contrary, you have one set of
: coefficients for each category compared to the reference. And that's what I
: want.
:
: ?
| F****n
发帖数: 3271 | 8 Actually the results are not bad. Technically it is not difficult. I just
feel a little nervous about the potential flaw in this approach.
Basically my dependent has 4 categories, and I have run the regression 3
times, each with a different reference category.
【在 l*********s 的大作中提到】
: Shall be fine. Plz proceed and let me know the result.
:
:
: need
: don
: not
: worry
|
|
