Doing it manually gives you more customization.

There is a typo in the syntax for the lines of the confidence limits. The model predictions
test.pred <- predict(test,newdata=data,se.fit=TRUE,type='response', na.action=na.omit)
are stored in test.pred.

Therefore the lines should be
lines( age[I1] , test.pred$fit[I1] + 1.96 * test.pred$se[I1] , lty = 2 )
Note that it is 1.96 times standard error for the 95% CI ( not 2 times). Ok, thats not so serious. But it has to be 1.96 * test.pred$se.

For the logistic regression model you would use
type = "link"
instead of
type = "response"
as far as I recall for the logit aka ln(OR).

Have a look at
?predict.gam

and remember
library(mgcv)
before you start.

Then both plots should be exactly the same.

Regards,
Reinhard

Reinhard

Of course it is possible. You might want to look at library(mgcv) of Simon Wood.

I wrote a post long ago about it, which I consider now outdated – one of my first approaches, but you might get some idea.
http://rforge.org/2009/08/11/gam-plot-with-95-confidence-shade/

a additive logistic regression model in mgcv would look like:
library(mgcv)
model <- gam( Event ~ s( Covariate1) + s(Covariate2) + factor(Cofactor1) , data = MyData , family = "binomial")
summary(model)
plot(model)
Note that family = “binomial” specifies the logit-link-function aka the logistic regression model.

Hope no typos, quite drafty written.

Regards, Reinhard

But I am very eager to know is there any similar programme to make a smoothing spline plot for a logistic model.

The relevant part is on page 108ff, chapter 4, Modeling the functional form of covariates.

Therneau is the maintainer of the “survival” package in R. I wonder that this plot method is not implemented more explicitly in “library(survival)”. As I mentioned in one of the additive survival posts, you can get some crude additive plots with the termplot() function…

Say you have a model
coxph( survival ~ pspline(cov1) + pspline (cov2) , data = MYDATA)

Then all observations with missing covariate for cov1 and cov2 are excluded. To mask those observations from the for plotting you might create an index variable which is TRUE when the covariates are not missing:
i.complete <- with( MYDATA , !is.na(cov1) & !is.na(cov2))

and then plot
lines( sm.spline( MYDATA$x[i.complete] , predict$fit ))

I know, that is very unelegant… using a smooth to plot the line is not a good programming practice, but I am quite new to R, too.

I have some improvements for the code. I will post them soon. It would be better to sort the MYDATA$x and predict$fit according to the x-values and then just plot a line. Something like:
i.x <- order( MYDATA$x )
lines(MYDATA$x[i.x] , predict$fit[i.x])

I posted it in a recent post concerning usual regression GAMs.

This would save the sm.spline() and you would not even need to load the library(pspline).

Hope that works and no typos…