R help - May 2010 - Visualizing binary response data?

Hi All,

I'm dealing with binary response data for the first time, and I'm
confused
about what kind of graphics I could explore in order to pick relevant
predictors and their relation with response variable.

I have 8-10 continuous predictors and 4-5 categorical predictors. Can anyone
suggest what kind of graphics I can explore to see how predictors behave
w.r.t. response variable...

Any help would be greatly appreciated, thanks,
Kim

	[[alternative HTML version deleted]]

For binary w.r.t. continuous, how about a smoothing spline?  As in,

x<-rnorm(100)
y<-rbinom(100,1,exp(.3*x-.07*x^2)/(1+exp(.3*x-.07*x^2)))
plot(x,y)
lines(smooth.spline(x,y))

OR how about a more parametric approach, logistic regression?  As in,

glm1<-glm(y~x+I(x^2),family=binomial)
plot(x,y)
lines(sort(x),predict(glm1,newdata=data.frame(x=sort(x)),type="response"))

FOR binary w.r.t. categorical it depends.  Are the categories ordinal (is
there a natural ordering?) or are the categories nominal (no ordering)?  For
nominal categories, the data is essentially a contingency table, and
"strength of the predictor" is a test of independence.  You can still
do a
graphical exploration: maybe plotting the proportion of Y=1 for each
category of X.   As in,

z<-cut(x,breaks=-3:3)
plot(tapply(y,z,mean))

If your goal is to find strong predictors of Y, you may want to consider
graphical measures that look at the predictors jointly.  Maybe with a
generalized additive model (gam)?

There is probably a lot more you can do.  Be creative.

-tgs



On Tue, May 4, 2010 at 9:04 PM, Kim Jung Hwa
<kimhwamaillist@gmail.com>wrote:
> Hi All,
>
> I'm dealing with binary response data for the first time, and I'm
confused
> about what kind of graphics I could explore in order to pick relevant
> predictors and their relation with response variable.
>
> I have 8-10 continuous predictors and 4-5 categorical predictors. Can
> anyone
> suggest what kind of graphics I can explore to see how predictors behave
> w.r.t. response variable...
>
> Any help would be greatly appreciated, thanks,
> Kim
>
>        [[alternative HTML version deleted]]
>
> ______________________________________________
> R-help@r-project.org mailing list
> https://stat.ethz.ch/mailman/listinfo/r-help
> PLEASE do read the posting guide
> http://www.R-project.org/posting-guide.html
> and provide commented, minimal, self-contained, reproducible code.
>
	[[alternative HTML version deleted]]

You could also try using interactive graphics in iplots.  Linking from a
barchart of your binary response variable to your eight continuous predictors in
a parallel coordinate plot and to your four categorical predictors in some form
of mosaicplot could be very informative.

Graphics are not necessarily the method of choice to select your predictor
variables, as Frank Harrell has pointed out.  It is also sensible not to rely on
modelling alone.  Graphic displays can help you better understand your data and
models.  The two approaches are complementary.

Antony Unwin
University of Augsburg
Germany


On Tue, May 4, 2010 at 9:04 PM, Kim Jung Hwa
<kimhwamaillist@gmail.com>wrote:
> Hi All,
> 
> I'm dealing with binary response data for the first time, and I'm
confused
> about what kind of graphics I could explore in order to pick relevant
> predictors and their relation with response variable.
> 
> I have 8-10 continuous predictors and 4-5 categorical predictors. Can
> anyone
> suggest what kind of graphics I can explore to see how predictors behave
> w.r.t. response variable...
> 
> Any help would be greatly appreciated, thanks,
> Kim
> ______________________________________________

	[[alternative HTML version deleted]]