| |
 |
|
| Science Forum Index » Space - Consult Forum » Back Transformations of Arcsine Square root... |
|
Page 1 of 1 |
|
| Author |
Message |
| DBK... |
 Posted: Wed Sep 09, 2009 4:29 pm |
|
|
|
Guest
|
Before running a standard logistic regression (and later a multi-level
logistic model), I applied the arcsine, square root transformation to
several variables expressed originally in proportions. I have two
questions. First, is there any reason I shouldn't do this? Second, if
not, what would be the back transformation procedure? First convert
from log of odds ratio to odds ratio? Then a back transformation of
the arcsine square root transformation?
DBK |
|
|
| Back to top |
|
|
|
| Bruce Weaver... |
| Posted: Thu Sep 10, 2009 1:09 am |
|
|
|
Guest
|
On Sep 9, 10:29 pm, DBK <boydkra... at (no spam) gmail.com> wrote:
[quote:7787ed7f49]Before running a standard logistic regression (and later a multi-level
logistic model), I applied the arcsine, square root transformation to
several variables expressed originally in proportions. I have two
questions. First, is there any reason I shouldn't do this? Second, if
not, what would be the back transformation procedure? First convert
from log of odds ratio to odds ratio? Then a back transformation of
the arcsine square root transformation?
DBK
[/quote:7787ed7f49]
In my experience, the arcsine transformation is typically used for
dependent variables that are proportions (usually in the context of
ANOVA or linear regression models). Why do you want to use it for
explanatory variables? In logistic regression, there is no
requirement for continuous explanatory variables to be normally
distributed, if that is what you are concerned about.
--
Bruce Weaver
bweaver at (no spam) lakeheadu.ca
http://sites.google.com/a/lakeheadu.ca/bweaver/Home
"When all else fails, RTFM." |
|
|
| Back to top |
|
|
|
| DBK... |
| Posted: Thu Sep 10, 2009 2:19 am |
|
|
|
Guest
|
On Sep 10, 7:09 am, Bruce Weaver <bwea... at (no spam) lakeheadu.ca> wrote:
[quote:ca08d4cf66]On Sep 9, 10:29 pm, DBK <boydkra... at (no spam) gmail.com> wrote:
Before running a standard logistic regression (and later a multi-level
logistic model), I applied the arcsine, square root transformation to
several variables expressed originally in proportions. I have two
questions. First, is there any reason I shouldn't do this? Second, if
not, what would be the back transformation procedure? First convert
from log of odds ratio to odds ratio? Then a back transformation of
the arcsine square root transformation?
DBK
In my experience, the arcsine transformation is typically used for
dependent variables that are proportions (usually in the context of
ANOVA or linear regression models). Why do you want to use it for
explanatory variables? In logistic regression, there is no
requirement for continuous explanatory variables to be normally
distributed, if that is what you are concerned about.
--
Bruce Weaver
bwea... at (no spam) lakeheadu.cahttp://sites.google.com/a/lakeheadu.ca/bweaver/Home
"When all else fails, RTFM."
[/quote:ca08d4cf66]
Yes, there is not a normality requirement for a standard logistic
model (i.e. single level) but what about a multilevel logistic model?
Thanks for the response. |
|
|
| Back to top |
|
|
|
| Bruce Weaver... |
| Posted: Thu Sep 10, 2009 3:39 am |
|
|
|
Guest
|
On Sep 10, 8:19 am, DBK <boydkra... at (no spam) gmail.com> wrote:
[quote:dd2be92f88]On Sep 10, 7:09 am, Bruce Weaver <bwea... at (no spam) lakeheadu.ca> wrote:
On Sep 9, 10:29 pm, DBK <boydkra... at (no spam) gmail.com> wrote:
Before running a standard logistic regression (and later a multi-level
logistic model), I applied the arcsine, square root transformation to
several variables expressed originally in proportions. I have two
questions. First, is there any reason I shouldn't do this? Second, if
not, what would be the back transformation procedure? First convert
from log of odds ratio to odds ratio? Then a back transformation of
the arcsine square root transformation?
DBK
In my experience, the arcsine transformation is typically used for
dependent variables that are proportions (usually in the context of
ANOVA or linear regression models). Why do you want to use it for
explanatory variables? In logistic regression, there is no
requirement for continuous explanatory variables to be normally
distributed, if that is what you are concerned about.
Yes, there is not a normality requirement for a standard logistic
model (i.e. single level) but what about a multilevel logistic model?
Thanks for the response.
[/quote:dd2be92f88]
I don't have any particular expertise in multilevel logistic
regression. But for multilevel linear regression (with two levels),
the level-1 residuals are assumed to be normal with a constant
variance; and the level-2 residuals are assumed to "have a
multivariate normal distribution with a constant covariance
matrix" (Snijders & Bosker, 1999, p. 121). To assess how well you are
meeting those assumptions, you need to run your model and look at the
residuals.
--
Bruce Weaver
bweaver at (no spam) lakeheadu.ca
http://sites.google.com/a/lakeheadu.ca/bweaver/Home
"When all else fails, RTFM." |
|
|
| Back to top |
|
|
|
| DBK... |
| Posted: Thu Sep 10, 2009 7:02 am |
|
|
|
Guest
|
On Sep 10, 9:39 am, Bruce Weaver <bwea... at (no spam) lakeheadu.ca> wrote:
[quote:2fa3a66cc5]On Sep 10, 8:19 am, DBK <boydkra... at (no spam) gmail.com> wrote:
On Sep 10, 7:09 am, Bruce Weaver <bwea... at (no spam) lakeheadu.ca> wrote:
On Sep 9, 10:29 pm, DBK <boydkra... at (no spam) gmail.com> wrote:
Before running a standard logistic regression (and later a multi-level
logistic model), I applied the arcsine, square root transformation to
several variables expressed originally in proportions. I have two
questions. First, is there any reason I shouldn't do this? Second, if
not, what would be the back transformation procedure? First convert
from log of odds ratio to odds ratio? Then a back transformation of
the arcsine square root transformation?
DBK
In my experience, the arcsine transformation is typically used for
dependent variables that are proportions (usually in the context of
ANOVA or linear regression models). Why do you want to use it for
explanatory variables? In logistic regression, there is no
requirement for continuous explanatory variables to be normally
distributed, if that is what you are concerned about.
Yes, there is not a normality requirement for a standard logistic
model (i.e. single level) but what about a multilevel logistic model?
Thanks for the response.
I don't have any particular expertise in multilevel logistic
regression. But for multilevel linear regression (with two levels),
the level-1 residuals are assumed to be normal with a constant
variance; and the level-2 residuals are assumed to "have a
multivariate normal distribution with a constant covariance
matrix" (Snijders & Bosker, 1999, p. 121). To assess how well you are
meeting those assumptions, you need to run your model and look at the
residuals.
--
Bruce Weaver
bwea... at (no spam) lakeheadu.cahttp://sites.google.com/a/lakeheadu.ca/bweaver/Home
"When all else fails, RTFM."
[/quote:2fa3a66cc5]
Thanks Bruce. But again, I am running logistic models not linear
models. |
|
|
| Back to top |
|
|
|
| Bruce Weaver... |
| Posted: Thu Sep 10, 2009 10:14 am |
|
|
|
Guest
|
On Sep 10, 1:02 pm, DBK <boydkra... at (no spam) gmail.com> wrote:
[quote:bec4d26e4f]On Sep 10, 9:39 am, Bruce Weaver <bwea... at (no spam) lakeheadu.ca> wrote:
On Sep 10, 8:19 am, DBK <boydkra... at (no spam) gmail.com> wrote:
On Sep 10, 7:09 am, Bruce Weaver <bwea... at (no spam) lakeheadu.ca> wrote:
On Sep 9, 10:29 pm, DBK <boydkra... at (no spam) gmail.com> wrote:
Before running a standard logistic regression (and later a multi-level
logistic model), I applied the arcsine, square root transformation to
several variables expressed originally in proportions. I have two
questions. First, is there any reason I shouldn't do this? Second, if
not, what would be the back transformation procedure? First convert
from log of odds ratio to odds ratio? Then a back transformation of
the arcsine square root transformation?
DBK
In my experience, the arcsine transformation is typically used for
dependent variables that are proportions (usually in the context of
ANOVA or linear regression models). Why do you want to use it for
explanatory variables? In logistic regression, there is no
requirement for continuous explanatory variables to be normally
distributed, if that is what you are concerned about.
Yes, there is not a normality requirement for a standard logistic
model (i.e. single level) but what about a multilevel logistic model?
Thanks for the response.
I don't have any particular expertise in multilevel logistic
regression. But for multilevel linear regression (with two levels),
the level-1 residuals are assumed to be normal with a constant
variance; and the level-2 residuals are assumed to "have a
multivariate normal distribution with a constant covariance
matrix" (Snijders & Bosker, 1999, p. 121). To assess how well you are
meeting those assumptions, you need to run your model and look at the
residuals.
--
Bruce Weaver
bwea... at (no spam) lakeheadu.cahttp://sites.google.com/a/lakeheadu.ca/bweaver/Home
"When all else fails, RTFM."
Thanks Bruce. But again, I am running logistic models not linear
models.
[/quote:bec4d26e4f]
What I meant to suggest was that if the assumptions for multilevel
linear models have to do with the residuals (rather than the variables
themselves), I should think that would be case for multilevel logistic
models too.
Also, it would seem quite odd (to me, at least) if multilevel logistic
models made assumptions about the distributions of explanatory
variables that are not required for ordinary (single-level) logistic
regression models. I can't lay my hands on any books or articles that
say this at the moment...but it would seem odd! Perhaps someone else
can jump in with a relevant book or article.
--
Bruce Weaver
bweaver at (no spam) lakeheadu.ca
http://sites.google.com/a/lakeheadu.ca/bweaver/Home
"When all else fails, RTFM." |
|
|
| Back to top |
|
|
|
| Ryan... |
| Posted: Fri Sep 11, 2009 4:19 am |
|
|
|
Guest
|
On Sep 10, 8:19 am, DBK <boydkra... at (no spam) gmail.com> wrote:
[quote:015233e1e3]On Sep 10, 7:09 am, Bruce Weaver <bwea... at (no spam) lakeheadu.ca> wrote:
On Sep 9, 10:29 pm, DBK <boydkra... at (no spam) gmail.com> wrote:
Before running a standard logistic regression (and later a multi-level
logistic model), I applied the arcsine, square root transformation to
several variables expressed originally in proportions. I have two
questions. First, is there any reason I shouldn't do this? Second, if
not, what would be the back transformation procedure? First convert
from log of odds ratio to odds ratio? Then a back transformation of
the arcsine square root transformation?
DBK
In my experience, the arcsine transformation is typically used for
dependent variables that are proportions (usually in the context of
ANOVA or linear regression models). Why do you want to use it for
explanatory variables? In logistic regression, there is no
requirement for continuous explanatory variables to be normally
distributed, if that is what you are concerned about.
--
Bruce Weaver
bwea... at (no spam) lakeheadu.cahttp://sites.google.com/a/lakeheadu.ca/bweaver/Home
"When all else fails, RTFM."
Yes, there is not a normality requirement for a standard logistic
model (i.e. single level) but what about a multilevel logistic model?
Thanks for the response.- Hide quoted text -
- Show quoted text -
[/quote:015233e1e3]
I'm cross-posting this question to the SAS group. The question is
around assumptions to running a multilevel logistic regression (a type
of generalized linear mixed model). Could someone explain and/or
provide a reference for the assumptions to running a multilevel
logistic regression model.
Let's stick with a simple example where we have a binary dependent
variable (0/1), one continuous explanatory variable, and a random
intercept. One might code up this type of model in the GLIMMIX
procedure as follows:
proc glimmix data=mydata method=quad;
class person;
model y = x / s link=logit dist=binary;
random intercept / subject = person;
run;
What are the specific assumptions to this test?
Thanks,
Ryan |
|
|
| Back to top |
|
|
|
|
|
All times are GMT - 5 Hours
The time now is Sun Nov 22, 2009 5:33 am
|
|