1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
|
\name{mkterm}
\alias{mkterm}
\title{
Assembling Model Terms for Smoothing Spline ANOVA Models
}
\description{
Assemble numerical functions for calculating model terms in a
smoothing spline ANOVA model.
}
\usage{
mkterm(mf, type)
}
\arguments{
\item{mf}{Model frame of the model formula.}
\item{type}{List specifying the type of spline for each variable.}
}
\section{Background}{
Tensor-product splines are constructed based on the model formula
and the marginal reproducing kernels, as described in Gu (2002,
Sec. 2.4). The marginal variables can be factors, numerical
vectors, and numerical matrices, as specified in the details
section.
One-way ANOVA decompositions are built in the supported marginal
constructions, in which one has the constant, a "nonparametric
contrast," and possibly also a "parametric contrast." To the
"nonparametric contrast" there corresponds a reproducing kernal
\code{rk}, and to a "parametric contrast" there corresponds a set
of null space basis \code{phi}. The reproducing kernels and null
space basis on the product domain can be constructed from the
marginal \code{rk} and \code{phi} in a systematic manner.
The marginal one-way ANOVA structures induce a multi-way ANOVA
structure on the product domain, with model terms consisting of
unpenalized "parametric contrasts" and/or penalized "nonparametric
contrasts." One only needs to construct \code{rk}'s and
\code{phi}'s associated with the model terms implied by the model
formula.
}
\details{
For a \bold{factor} \code{x}, \code{type$x} is ignored;
\code{\link{mkrk.ordinal}} is used if \code{is.ordered(x)==TRUE} and
\code{\link{mkrk.nominal}} is used otherwise. Factors with 3 or
more levels are penalized.
For a \bold{numerical vector} \code{x}, \code{type$x} is of the form
\code{type.x} for \code{type.x}=\code{"cubic"}, \code{"linear"},
or of the form \code{list(type.x, range)} for
\code{type.x}=\code{"per"}, \code{"cubic.per"}, \code{"linear.per"},
\code{"cubic"}, \code{"linear"}; \code{"per"} is short for
\code{"cubic.per"}. See \code{\link{mkfun.poly}} for the functions
used. For \code{type.x} missing, the \bold{default} is \code{"cubic"}.
For \code{range} missing with \code{type.x}=\code{"cubic"},
\code{"linear"}, the \bold{default} is
\code{c(min(x),max(x))+c(-1,1)*(max(x)-mimn(x))*.05}.
For a \bold{numerical matrix} \code{x}, \code{type$x} is of the form
\code{type.x} or \code{list(type.x, order)} for
\code{type.x}=\code{"tp"}, \code{"sphere"}, or of the form
\code{list("tp",list(order=order,mesh=mesh,weight=weight))}. See
\code{\link{mkfun.tp}} for the functions used. For \code{type.x}
missing, the \bold{default} is \code{"tp"}. For \code{order}
missing, the \bold{default} is \code{2}. For \code{mesh} and
\code{weight} missing with \code{type.x}=\code{"tp"} and
\code{order} given, the \bold{defaults} are \code{mesh}=\code{x} and
\code{weight}=\code{1}.
For a \bold{numerical vector} or \bold{numerical matrix} \code{x},
one may also use \code{type$x} of the form
\code{list("custom",list(nphi=nphi,mkphi=mkphi,mkrk=mkrk,env=env))};
\code{nphi} is the null space dimension \emph{excluding the
constant}, and \code{mkphi} is ignored if \code{nphi}=0. See
examples below. This feature allows the use of other marginal
constructions; one may modify \code{\link{mkphi.cubic}} or
\code{\link{mkphi.tp.p}} for \code{mkphi} and modify
\code{\link{mkrk.cubic}} or \code{\link{mkrk.sphere}} for
\code{mkrk}.
}
\note{
For a \bold{numerical vector} \code{x} in \code{\link{ssden}},
the default \code{range} is \code{domain$x}.
For a \bold{numerical matrix} \code{x} with
\code{type.x}=\code{"sphere"}, it is assumed that
\code{dim(x)[2]==2}, \code{x[,1]} between [-90,90] the latitude in
degrees, and \code{x[,2]} between [-180,180] the longitude in
degrees.
For \bold{backward compatibility}, one may set \code{type="cubic"},
\code{"linear"}, or \code{"tp"}, but then the default parameters can
not be overridden; the type is simply duplicated for each variable.
}
\value{
A list object with a component \code{labels} containing the labels
of all model terms. For each of the model terms, there is a
component holding the numerical functions for calculating the
unpenalized and penalized parts within the term.
}
\author{Chong Gu, \email{chong@stat.purdue.edu}}
\references{
Gu, C. (2013), \emph{Smoothing Spline ANOVA Models (2nd Ed)}. New
York: Springer-Verlag.
}
\examples{
## cubic marginals
x1 <- rnorm(100); x2 <- rnorm(100); y <- 3+5*sin(x1-2*x2)+rnorm(x1)
fit <- ssanova0(y~x1*x2)
## the same fit
fit1 <- ssanova0(y~x1*x2,type=list(x1="cubic"))
## the same fit one more time
par <- list(nphi=1,mkphi=mkphi.cubic,mkrk=mkrk.cubic,
env=c(min(x2),max(x2))+c(-1,1)*(max(x2)-min(x2))*.05)
fit2 <- ssanova0(y~x1*x2,type=list(x2=list("custom",par)))
## Clean up
\dontrun{rm(x1,x2,y,fit,fit1,par,fit2)}
## cubic and thin-plate marginals
x1 <- rnorm(100); x2 <- matrix(rnorm(200),100,2)
y <- 3+5*sin(x1-2*x2[,1]*x2[,2])+rnorm(x1)
fit <- ssanova0(y~x1*x2)
## the same fit
fit1 <- ssanova0(y~x1*x2,type=list(x2="tp"))
## the same fit one more time
mkphi.tp1 <- function(x) mkphi.tp(x$dm,x$ord,x$mesh,x$wt)
mkrk.tp1 <- function(x) mkrk.tp(x$dm,x$ord,x$mesh,x$wt)
env <- list(dm=2,ord=2,mesh=x2,wt=1)
par <- list(nphi=2,mkphi=mkphi.tp1,mkrk=mkrk.tp1,env=env)
fit2 <- ssanova0(y~x1*x2,type=list(x2=list("custom",par)))
## Clean up
\dontrun{rm(x1,x2,y,fit,fit1,mkphi.tp1,mkrk.tp1,env,par,fit2)}
}
\keyword{internal}
|
