Hello R-Users,
I have recently run into several problems using vgam() in the VGAM
package. I am hoping someone might have some solutions...
Briefly, I have been trying to fit GAM models for zero-altered negative
binomial models.
1. When fitting smoothed parameters (e.g. s(X, df=2)) changing the
degrees-of-freedom has no effect on the level of smoothing (e.g. number
of knots for the spline). This appears to be the case even for the
example on the vgam help page.
data(hunua)
fit2 = vgam(agaaus ~ s(altitude, df=2), binomialff, hunua)
> coef(fit2)
(Intercept) s(altitude, df = 2)
-1.1661259280 0.0003932463
fit10 = vgam(agaaus ~ s(altitude, df=10), binomialff, hunua)
> coef(fit10)
(Intercept) s(altitude, df = 10)
-1.1661259280 0.0003932463
2. There may be a bug when trying to run gam (mgcv) and vgam(VGAM)
together. When I run vgam first, I am unable to get smoothed parameters
using gam. This presents a problem if one would like to use the
parameter estimates from a gam model to try to fit a vgam model.
3. For my own data, I have been having a very difficult time getting
vgam to converge. The help pages says that convergence may depend
strongly on the initial values (coefstart) provided, and I have been
modifying them but to no avail. Is there another way or more values
that need to be given to make vgam models fit?
I frequently run into the error below. Any advice on how to overcome it?
Applying Greenstadt modification to 357 matrices
Error in ans[, index] <- tmp777 :
number of items to replace is not a multiple of replacement length
In addition: Warning message:
In checkwz(wz, M = M, trace = trace, wzeps = control$wzepsilon) :
357 elements replaced by 1.819e-12
In case someone has the interest or time in helping me get models to
converge, I provide an example model and data. (Data are attached and
the model code is below.) Thanks for your help.
library(VGAM)
Cnictdat=read.csv("Cnictdat.csv", header=T, sep=",")
coefnict=c(-6.249e+00, -8.049e+00, -1.0e+01, -1.66e+00,
6.620960e-01, -1.046e-02, 4.586179e-01, -1.189e+00,
-2.712e-03, -2.126218e+00, 7.242e-01, -1.149e+00,
9.377e-01, -8.457e-03, -3.035807e-02, -1.697e-02,
2.279305e-02, -4.7e-07, -1.797e-07, -1.183e-07, -4.3494e-05)
abunddst=c(Cnictdat$SegDistKm)*1000
esw=0.5
fit4=vgam(Cnict~factor(Hab)+s(Clr, spar=2)+s(Rd, spar=2)+s(YL,
spar=2)+ s(Pop, spar=2)+offset(log(2*esw*abunddst)),
zanegbinomial, data=Cnictdat, coefstart=coefnict,
control=vgam.control(se.fit=T, exp=1e-4, maxit=500,
trace=T))
Thanks,
John
