[R] Plotting GAM fit using RGL

[David Winsemius]

On Aug 16, 2013, at 10:55 AM, Duncan Murdoch wrote:

> On 13-08-15 1:15 PM, David Winsemius wrote:
>> 
>> On Aug 15, 2013, at 2:23 AM, Lucas Holland wrote:
>> 
>>> Hello all,
>>> 
>>> I’ve fitted a bivariate smoothing model (with GAM) to some data, using two explanatory variables, x and y.  Now I’d like to add the surface corresponding to my fit to a 3D scatterplot generated using plot3d().
>>> 
>>> My approach so far is to create a grid of x and y values and the corresponding predicted values and to try to use surface3d with that grid.
>>> 
>>> grid <- expand.grid(x = seq(-1,1,length=20),
>>>                    y = seq(-1,1, length=20))
>>> 
>>> grid$z <- predict(fit.nonparametric, newdata=grid)
>>> 
>>> surface3d(grid$x, grid$y, matrix(grid$z, nrow=length(grid$x), ncol=length(grid$y)))
>>> 
>> ?surface3d
>> # Should be:
>> 
>>  surface3d( unique(grid$x), unique(grid$y),
>>                     z= matrix(grid$z, nrow=length(grid$x), ncol=length(grid$y)))
> 
> Or you could make x and y into matrices as well.  In this case you'll get the same result, but if x or y weren't strictly increasing sequences, there'd be a difference.

Thanks for increasing my knowledge on this point. And for providing rgl to the world. After looking at the Details section of the help page more carefully than I had previously, I wondered: Has anyone ever done a projection of a Klein bottle into rgl?

I didn't find one and my initial efforts with surface3d failed. (I managed to crash that seesion with a misguided call to the global replace function.)

I did get success with misc3d's parameteric3d with a parametrisation attributed to Robert Israel:

require(rgl); require(misc3d)

x = function(u,v){-(2/15)*cos(u)*(3*cos(v)-30*sin(u)+90*cos(u)^4*sin(u)- 60*cos(u)^6*sin(u)+5*cos(u)*cos(v)*sin(u))}

y = function(u,v){-(1/15)*sin(u)*(3*cos(v)-3*cos(u)^2*cos(v)-48*cos(u)^4*cos(v)+48*cos(u)^6*cos(v)-60*sin(u)+5*cos(u)*cos(v)*sin(u) 
    -5*cos(u)^3*cos(v)*sin(u) -80*cos(u)^5*cos(v)*sin(u)+80*cos(u)^7*cos(v)*sin(u))}

z = function(u,v){ (2/15)*(3+5*cos(u)*sin(u))*sin(v) }

parametric3d(x,y,z, seq(0,pi,length=100), seq(0,2*pi,length=100) )

-- 
David.

> Duncan Murdoch
> 
>> 
>> 
>>> This however plots a number of surfaces that do not look like the fitted surface obtained by vis.gam(fit.nonparametric which actually looks a lot like the „truth“ (I’m using simulated data so I know the true regression surface).
>>> 
>>> I think I’m using surface3d wrong but I can’t seem to spot my mistake.
>> 
>> 
>> Always look at the Arguments section of help pages carefully.
>> 
> 

David Winsemius
Alameda, CA, USA