[R] package mgcv - predict with bam: Error in X[ind, ] : subscript out of bounds

[Simon Wood]

> I suppose there may be an error of sorts,  but have you considered
> the fact that solving the error might not gain you admittance into
> heaven? Look at the RHS of the model:
>
> sensor2 + s(site, bs = "re")
>
> ... and think about the fact that you are "smoothing" a factor
> variable.
- Actually this is ok. mgcv exploits the duality between quadratically
penalized smooths and Gaussian random effects to allow random effects to
be specified this way. bs="re" specifies a Gaussian random effect with 
corresponding model matrix given by model.matrix(~site-1). (More 
generally s(x,y,z,bs="re") specifies a gaussian random effect with model 
matrix given by model.matrix(~x:y:z-1), with obvious generalization to 
more or fewer variables). See mgcv help file ?random.effects for more.

best,
Simon



>> str(gapData)
> 'data.frame':	2304 obs. of  5 variables: $ sensor1    : num  NA NA NA
> NA NA NA NA NA NA NA ... $ site       : Factor w/ 9 levels
> "KRB","NP.FOR",..: 3 3 3 3 3 3 3 3 3 3 ... $ NthSampling: int  7489
> 7490 7491 7492 7493 7494 7495 7496 7497 7498 ... $ YDay       : num
> 53 53 53 53 53 53 53 53 53 53 ... $ sensor2    : num  0.567 0.566
> 0.567 0.567 0.569 ...
>
> I'm having trouble making any sense of how that might work. It is, of
> course, possible to just do this:
>
> xylemRohWeekXnnSite.fit <- predict.gam(xylemRohWeekXnn.fit.bam,
> type="response", se=F)
>
> That gives predictions for the original dataset.
>
> But I think the error might be helpful in alerting one to the
> problems with the model.
>


-- 
Simon Wood, Mathematical Science, University of Bath BA2 7AY UK
+44 (0)1225 386603               http://people.bath.ac.uk/sw283