blapsr (version 0.6.1)

penaltyplot: Plot the approximate posterior distribution of the penalty vector.

Description

The routine gives a graphical representation of the univariate approximate posterior distribution of the (log-)penalty parameters for objects of class coxlps, curelps, amlps and gamlps.

Usage

penaltyplot(object, dimension, ...)

Arguments

object

An object of class coxlps, curelps, amlps or gamlps.

dimension

For objects of class amlps and gamlps, the penalty vector can have a dimension larger than one, i.e. more than a single smooth term is present in the considered additive model. In that case, dimension is the penalty dimension to be plotted corresponding either to a scalar indicating the desired dimension or to a vector indicating more than one dimension. For instance, dimension = c(1,3) displays two separate plots of the (approximate) posterior distribution of the (log-)penalty parameter associated to the first and the third smooth function respectively.

...

Further arguments to be passed to the routine.

Author

Oswaldo Gressani .

Details

When q, the number of smooth term in a (generalized) additive model is smaller than five, the exploration of the posterior penalty space is based on a grid strategy. In particular, the multivariate grid of dimension q is constructed by taking the Cartesian product of univariate grids in each dimension j = 1,...q. These univariate grids are obtained from a skew-normal fit to the conditional posterior p(vj|vmap[-j]),D), where vj is the (log-)penalty value associated to the jth smooth function and vmap[-j] is the posterior maximum of the (log-)penalty vector omitting the jth dimension. The routine displays the latter skew-normal distributions. When q>=5, inference is based on vmap and the grid is omitted to avoid computational overflow. In that case, the posterior distribution of the (log-)posterior penalty vector v is approximated by a multivariate Gaussian and the routine shows the marginal distributions.

Examples

### Classic simulated data example (with simgamdata)

set.seed(123)
sim.data <- simgamdata(setting = 2, n = 250, dist = "gaussian", scale = 0.25)
plot(sim.data)         # Scatter plot of response
data <- sim.data$data  # Simulated data frame
# Fit model
fit <- amlps(y ~ z1 + z2 + sm(x1) + sm(x2), data = data, K = 15)
fit

# Penalty plot
opar <- par(no.readonly = TRUE)
par(mfrow = c(1, 2))
penaltyplot(fit, dimension = c(1, 2))
par(opar)