vcrpart (version 1.0-6)

fvcm-methods: Methods for fvcm objects

Description

Standard methods for computing on fvcm objects.

Usage

# S3 method for fvcm
oobloss(object, fun = NULL, ranef = FALSE, ...)
# S3 method for fvcm
plot(x, type = c("default", "coef", 
           "simple", "partdep"),
     tree = NULL, ask = NULL, ...)
# S3 method for fvcm
predict(object, newdata = NULL,
        type = c("link", "response", "prob", "class", "coef", "ranef"),
        ranef = FALSE, na.action = na.pass, verbose = FALSE, ...)

Value

The methods fitted.fvcm and

predict.fvcm return an object of class numeric

or matrix, depending on the used model or the specification of the argument type. See also fitted.tvcm.

The oobloss.fvcm method returns the output of the loss function defined by fun. This is a single numeric by default. See also oobloss.

The plot.fvcm method returns NULL.

The ranef.fvcm method returns an object of class

matrix with values for the random effects. See also

ranef.olmm and ranef.

Arguments

object, x

an object of class fvcm.

fun

the loss function. The default loss function is defined as the sum of the deviance residuals. For a user defined function fun, see the examples of oobloss.tvcm.

newdata

an optional data frame in which to look for variables with which to predict. If omitted, the training data are used.

type

character string indicating the type of plot or prediction. See plot.tvcm or predict.tvcm. "response" and "prob" are identical.

tree

integer vector. Which trees should be plotted.

ask

logical. Whether an input should be asked before printing the next panel.

ranef

logical scalar or matrix indicating whether predictions should be based on random effects. See predict.olmm.

na.action

function determining what should be done with missing values for fixed effects in newdata. The default is to predict NA: see na.pass.

verbose

logical scalar. If TRUE verbose output is generated during the validation.

...

further arguments passed to other methods.

Author

Reto Burgin

Details

oobloss.fvcm estimates the out-of-bag loss based on predictions of the model that aggregates only those trees in which the observation didn't appear (cf. Hastie et al, 2001, sec. 15). The prediction error is computed as the sum of prediction errors obtained with fun, which are the deviance residuals by default.

The plot and the prediction methods are analogous to plot.tvcm resp. predict.tvcm. Note that the plot options mean and conf.int for type ="coef" are not available (and internally set to FALSE).

Further undocumented, available methods are fitted, print and ranef. All these latter methods have the same arguments as the corresponding default methods.

References

Breiman, L. (1996). Bagging Predictors. Machine Learning, 24(2), 123--140.

Breiman, L. (2001). Random Forests. Machine Learning, 45(1), 5--32.

Hastie, T., R. Tibshirani and J. Friedman (2001). The Elements of Statistical Learning (2 ed.). New York, USA: Springer-Verlag.

See Also

fvcm, tvcm-methods

Examples

## ------------------------------------------------------------------- #
## Dummy example 1:
##
## Fitting a random forest tvcm on artificially generated ordinal
## longitudinal data. The parameters 'maxstep = 1' and 'K = 2' are     
## chosen to restrict the computations.
## ------------------------------------------------------------------- # 

## load the data

data(vcrpart_1)

## fit and analyse the model

control <-
  fvcolmm_control(mtry = 2, maxstep = 1, 
                  folds = folds_control(type = "subsampling", K = 2, prob = 0.75))

model.1 <-
  fvcolmm(y ~ -1 + wave + vc(z3, z4, by = treat, intercept = TRUE) + re(1|id),
          family = cumulative(), subset = 1:100,
          data = vcrpart_1, control = control)

## estimating the out of bag loss
suppressWarnings(oobloss(model.1))

## predicting responses and varying coefficients for subject '27'
subs <- vcrpart_1$id == "27"

## predict coefficients
predict(model.1, newdata = vcrpart_1[subs,], type = "coef")

## marginal response prediction
predict(model.1, vcrpart_1[subs,], "response", ranef = FALSE)

## conditional response prediction
re <- matrix(5, 1, 1, dimnames = list("27", "(Intercept)"))
predict(model.1, vcrpart_1[subs,], "response", ranef = re)
predict(model.1, vcrpart_1[subs,], "response", ranef = 0 * re)

## predicting in-sample random effects
head(predict(model.1, type = "ranef"))

## fitted responses (marginal and conditional prediction)
head(predict(model.1, type = "response", ranef = FALSE))
head(predict(model.1, type = "response", ranef = TRUE))


## ------------------------------------------------------------------- #
## Dummy example 2:
##
## Fitting a random forest tvcm on artificially generated normally
## distributed data. The parameters 'maxstep = 3' and 'K = 3' are
## chosen to restrict the computations and 'minsize = 5' to obtain at
## least a few splits given the small sample size.
## ------------------------------------------------------------------- #

data(vcrpart_2)

## fit and analyse the model

control <- fvcm_control(mtry = 1L, minsize = 5, maxstep = 3,
                        folds_control("subsampling", K = 3, 0.75))

model.2 <- fvcglm(y ~ -1  + vc(z1, z2, by = x1, intercept = TRUE) + x2,
                  data = vcrpart_2,
                  family = gaussian(), subset = 1:50,control = control)

## estimating the out of bag loss
suppressWarnings(oobloss(model.2))

## predict the coefficient for individual cases
predict(model.2, vcrpart_2[91:100, ], "coef")