R help - Jan 2006 - Binary logistic modelling: setting conditions (defining thresholds) in the fitted model (lrm)

Dear Rlist,

We are working with library(Design) & R 2.2.1//
When using the following fitted model:
	knots  <- 5
	lrm.1        <- lrm(X8~rcs(X1,5),x=T,y=T)

X8 (binary 0/1 vector)
X1, X2 explantory variables

We would like to set the probability of X8=1 to zero when the X2
variable is smaller than a defined threshold,
e.g. X2<50, because the X1 variable is not correct (contains more
errors) anymore when X2<50.

How could we  define this in the model smoothly without changing the
values of the variables? 

We keep in mind that setting thresholds in not a good solution because
then information is lost. Therefore we also tested the following model.
However, towards operational methods or techniques setting thresholds is
simplifying relationships. Especially in this case were we saw that X1
could contain more errors when X2 < 50.

lrm.1        <- lrm(X8~rcs(X1,5)+ rcs(X2,5),x=T,y=T)

Thanks a lot for feedback & discussion,
Jan




Disclaimer: http://www.kuleuven.be/cwis/email_disclaimer.htm

Jan Verbesselt wrote:
> Dear Rlist,
> 
> We are working with library(Design) & R 2.2.1//
> When using the following fitted model:
> 	knots  <- 5
> 	lrm.1        <- lrm(X8~rcs(X1,5),x=T,y=T)
> 
> X8 (binary 0/1 vector)
> X1, X2 explantory variables
> 
> We would like to set the probability of X8=1 to zero when the X2
> variable is smaller than a defined threshold,
> e.g. X2<50, because the X1 variable is not correct (contains more
> errors) anymore when X2<50.
Are you sure you want the prob(X8=1) to be zero or to you want to just 
constrain the regression function to be of a certain form?  And keep in 
mind that if the measurement errors are moderate or better it is usually 
  better to use the variable in its original form because otherwise real 
predictive information is lost.

Frank
> 
> How could we  define this in the model smoothly without changing the
> values of the variables? 
> 
> We keep in mind that setting thresholds in not a good solution because
> then information is lost. Therefore we also tested the following model.
> However, towards operational methods or techniques setting thresholds is
> simplifying relationships. Especially in this case were we saw that X1
> could contain more errors when X2 < 50.
> 
> lrm.1        <- lrm(X8~rcs(X1,5)+ rcs(X2,5),x=T,y=T)
> 
> Thanks a lot for feedback & discussion,
> Jan


-- 
Frank E Harrell Jr   Professor and Chair           School of Medicine
                      Department of Biostatistics   Vanderbilt University