Construct BayesX Model Term Objects
The function bayesx.construct is used to provide a flexible framework to implement
new model term objects in bayesx within the BayesX syntax.
bayesx.construct(object, dir, prg, data)
object |
is a smooth, shrinkage or random specification object in a STAR |
dir |
|
prg |
if additional data handling must be applied, e.g. storing maps ( |
data |
if additional data is needed to setup the BayesX term it is found here. |
The main idea of these constructor functions is to provide a flexible framework to implement new
model term objects in the BayesX syntax within bayesx, i.e. for any smooth or
random term in R2BayesX a constructor function like
bayesx.construct.ps.smooth.construct may be provided to translate R specific syntax into
BayesX readable commands. During processing with write.bayesx.input each model
term is constructed with bayesx.construct after another, wrapped into a full formula, which
may then be send to the BayesX binary with function run.bayesx.
At the moment the following model terms are implemented:
"rw1", "rw2": Zero degree P-splines: Defines a zero degree P-spline with first or
second order difference penalty. A zero degree P-spline typically
estimates for every distinct covariate value in the dataset a separate
parameter. Usually there is no reason to prefer zero degree P-splines
over higher order P-splines. An exception are ordinal covariates or
continuous covariates with only a small number of different values.
For ordinal covariates higher order P-splines are not meaningful while
zero degree P-splines might be an alternative to modeling nonlinear
relationships via a dummy approach with completely unrestricted
regression parameters.
"season": Seasonal effect of a time scale.
"ps", "psplinerw1", "psplinerw2": P-spline with first or second order
difference penalty.
"te", "pspline2dimrw1": Defines a two-dimensional P-spline based on the tensor
product of one-dimensional P-splines with a two-dimensional first order random walk
penalty for the parameters of the spline.
"kr", "kriging": Kriging with stationary Gaussian random fields.
"gk", "geokriging": Geokriging with stationary Gaussian random fields: Estimation
is based on the centroids of a map object provided in
boundary format (see function read.bnd and shp2bnd) as an additional
argument named map within function sx, or supplied within argument
xt when using function s, e.g., xt = list(map = MapBnd).
"gs", "geospline": Geosplines based on two-dimensional P-splines with a
two-dimensional first order random walk penalty for the parameters of the spline.
Estimation is based on the coordinates of the centroids of the regions
of a map object provided in boundary format (see function read.bnd and
shp2bnd) as an additional argument named map (see above).
"mrf", "spatial": Markov random fields: Defines a Markov random field prior for a
spatial covariate, where geographical information is provided by a map object in
boundary or graph file format (see function read.bnd, read.gra and
shp2bnd), as an additional argument named map (see above).
"bl", "baseline": Nonlinear baseline effect in hazard regression or multi-state
models: Defines a P-spline with second order random walk penalty for the parameters of
the spline for the log-baseline effect log(λ(time)).
"factor": Special BayesX specifier for factors, especially meaningful if
method = "STEP", since the factor term is then treated as a full term,
which is either included or removed from the model.
"ridge", "lasso", "nigmix": Shrinkage of fixed effects: defines a
shrinkage-prior for the corresponding parameters
γ_j, j = 1, …, q, q ≥q 1 of the
linear effects x_1, …, x_q. There are three
priors possible: ridge-, lasso- and Normal Mixture
of inverse Gamma prior.
"re": Gaussian i.i.d.\ Random effects of a unit or cluster identification covariate.
The model term syntax used within BayesX as a character string.
If new bayesx.construct functions are implemented in future work, there may occur problems
with reading the corresponding BayesX output files with read.bayesx.output,
e.g., if the new objects do not have the structure as implemented with bs = "ps" etc.,
i.e. function read.bayesx.output must also be adapted in such cases.
Using sx additional controlling arguments may be supplied within the dot dot dot
“...” argument. Please see the help site for function bayesx.term.options
for a detailed description of possible optional parameters.
Within the xt argument in function s, additional
BayesX specific parameters may be also supplied, see the examples below.
Nikolaus Umlauf, Thomas Kneib, Stefan Lang, Achim Zeileis.
bayesx.construct(sx(x1, bs = "ps"))
bayesx.construct(sx(x1, x2, bs = "te"))
## now create BayesX syntax for smooth terms
## using mgcv constructor functions
bayesx.construct(s(x1, bs = "ps"))
## for tensor product P-splines,
bayesx.construct(s(x1, x2, bs = "te"))
## increase number of knots
## for a P-spline
bayesx.construct(sx(x1, bs = "ps", nrknots = 40))
## now with degree 2 and
## penalty order 1
bayesx.construct(sx(x1, bs = "ps", knots = 40, degree = 2, order = 1))
bayesx.construct(s(x1, bs = "ps", k = 41, m = c(0, 1)))
## random walks
bayesx.construct(sx(x1, bs = "rw1"))
bayesx.construct(sx(x1, bs = "rw2"))
## shrinkage priors
bayesx.construct(sx(x1, bs = "lasso"))
bayesx.construct(sx(x1, bs = "ridge"))
bayesx.construct(sx(x1, bs = "nigmix"))
## for cox models, baseline
bayesx.construct(sx(time, bs = "bl"))
## kriging
bayesx.construct(sx(x, z, bs = "kr"))
## seasonal
bayesx.construct(sx(x, bs = "season"))
## factors
bayesx.construct(sx(id, bs = "factor"))
## now with some geographical information
## note: maps must be either supplied in
## 'bnd' or 'gra' format, also see function
## read.bnd() or read.gra()
data("MunichBnd")
bayesx.construct(sx(id, bs = "mrf", map = MunichBnd))
## same with
bayesx.construct(s(id, bs = "mrf", xt = list(map = MunichBnd)))
bayesx.construct(sx(id, bs = "gk", map = MunichBnd))
bayesx.construct(sx(id, bs = "gs", map = MunichBnd))
## also vary number of knots
bayesx.construct(sx(id, bs = "gs", knots = 10, map = MunichBnd))
bayesx.construct(s(id, bs = "gs", k = 12, m = c(1, 1), xt = list(map = MunichBnd)))
## random effects
bayesx.construct(sx(id, bs = "re"))
bayesx.construct(sx(id, bs = "re", by = x1))
bayesx.construct(sx(id, bs = "re", by = x1, xt = list(nofixed=TRUE)))
## generic
## specifies some model term
## and sets all additional arguments
## within argument xt
## only for experimental use
bayesx.construct(sx(x, bs = "generic", dosomething = TRUE, a = 1, b = 2))Please choose more modern alternatives, such as Google Chrome or Mozilla Firefox.