gam / bam / bam（discrete = TRUE）不包含/不包含/不包含交互（predict（type =“ terms”））

Question

When using the bam(discrete=TRUE) option, the predict(type="terms") function no longer reports interaction terms. This can be avoided by putting interaction terms in as separate, interacted columns, preventing the function from seeing the source of the interaction, but this is troublesome. Is there some theoretical reason we should not expect to see interactions from bam(discrete=TRUE), even though we can accomplish the same thing by creating the interactions explicitly?

The problem is fixable whenever we replace the interaction terms x:y with a bunch of columns explicitly encoding the interaction. I have tried other, various fixes in an attempt to force the interactions through; eg (0+x):y, which I'm not sure is even meaningful, suspecting that the columns are removed in the same way the intercept is removed from predict.bam(type="terms").

library(mgcv)

mydf <- data.frame(class = factor(round(runif(500), digits=1)),
               x     = runif(500),
               y     = runif(500))

mydf$z <- 1 + as.numeric(mydf$class) * mydf$x + 6*(mydf$y - 0.5)^2 + 
rnorm(n=500, mean=0, sd=0.1)

mygam <- gam(z ~ 0 + class + class:x + s(y), data=mydf)

head(predict(mygam, type="terms"))

mybam <- bam(z ~ 0 + class + class:x + s(y), data=mydf)

head(predict(mybam, type="terms"))

mydiscretebam <- bam(z ~ 0 + class + class:x + s(y), discrete=TRUE, data=mydf)

head(predict(mydiscretebam, type="terms"))

I expect the results to differ, since there are two different optimization algorithms in use, but I'm surprised that the interactions are not even reported in the third predict call. The model estimates in this case are not the sum of the terms. Why is the intercept removed? Is this also responsible for the interactions being removed? The coefficients are available in the summary, same as any other parametric coefficient, so is their exclusion accidental?

As far as I know, I'm really just changing the optimization algorithm. The fact that certain columns come from interactions seems irrelevant.

Answer 1

It's a bug in predict.gam (which is called by the discrete version of predict.bam to do the parametric terms). Fixed for the next release. Thanks. Meanwhile you can use

predict(mydiscretebam, type="terms",discrete=FALSE)

Simon (mgcv maintainer)