Skip to content
Prev 285239 / 398503 Next

mgcv: increasing basis dimension

That's interesting. Playing with the example, it doesn't seem to be a 
local minimum. I think that this happens because, although the higher 
rank basis contains the lower rank basis, the penalty can not simply 
suppress all the extra components in the higher rank basis and recover 
exactly what the lower rank basis gave: it's forced to include some of 
the extra stuff, even if heavily penalized, and this is what is 
degrading the higher rank fit in this case.

t2 tensor product smooths seem to be less susceptible to this effect, 
and for reasons I don't understand so does REML based smoothness 
selection (gam(...,method="REML"))

best,
Simon
On 13/02/12 23:24, Greg Dropkin wrote: