Construct BayesX Model Term Objects

Share:

Description

The function bayesx.construct is used to provide a flexible framework to implement new model term objects in bayesx within the BayesX syntax.

Usage

1
bayesx.construct(object, dir, prg, data)

Arguments

object

is a smooth, shrinkage or random specification object in a STAR formula, generated by the R2BayesX model term constructor functions sx (or using the constructor functions s and te of the mgcv package). Objects generated by these functions have class "xx.smooth.spec" where "xx" is determined by the "bs" argument of sx (and s).

dir

character, a directory where possible data should be stored, e.g. in bayesx models, if bs = "gk", bs = "gs" or bs = "mrf" is choosen, the corresponding map will be written as a "bnd" or "gra" file (see read.bnd and read.gra) to this directory, so BayesX can use this spatial object for estimation.

prg

if additional data handling must be applied, e.g. storing maps ("bnd") objects in the directory specified in dir, write.bayesx.input needs to write the extra commands in a program file provided with argument prg, i.e. this may all be handled within a bayesx.construct constructor function.

data

if additional data is needed to setup the BayesX term it is found here.

Details

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.

See function sx for a description of the main R2BayesX model term constructor functions.

Value

The model term syntax used within BayesX as a character string.

WARNINGS

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.

Note

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.

Author(s)

Nikolaus Umlauf, Thomas Kneib, Stefan Lang, Achim Zeileis.

See Also

sx, bayesx.term.options, s, formula.gam, read.bnd, read.gra.

Examples

 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
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))

Want to suggest features or report bugs for rdrr.io? Use the GitHub issue tracker.