MRSP-class: Class '"MRSP"'

In MRSP: Multinomial Response Models with Structured Penalties

Description

An object containing the output of fitting a multinomial response model with structured penalties for one concrete value of the tuning parameter (or one concrete combination of all used tuning parameters).

Objects from the Class

Objects of this class are created by function MRSP.fit, which will most likely be called internally by the end-user function MRSP.

Slots

coef:

Object of class "MRSP.coef" containing the estimated regression coefficients. Structurally, it is simply a list of one or two matrices.

coef.stand:

Same as slot coef, but contains the coefficients belonging to standardized covariates.

coef.pretres:

Same as slot coef, but contains the coefficients belonging to standardized covariates before any (potential) thresholding took place.

dat:

Data object in the form required by MRSP.fit. To save memory, this slot is usually set to NULL.

x.original:

Original matrix containing the "x"-variables. To save memory, this slot is often times set to NULL.

x.stand:

The "x"-matrix after standardization.

V.original:

The original data object containing the category-specific predictors. To save memory, this slot is often times NULL.

V.stand:

The "V"-object after standardization.

y:

Response matrix in the form required by MRSP.fit. To save memory, this slot is often times set to NULL.

weights:

Vector of observation weights.

penindex:

Object specifying how each covariate is penalized. See the documentation of MRSP.fit for details.

grpindex:

Object specifying which predictors form parameter groups. See the documentation of MRSP.fit for details.

penweights:

Object specifying the weighting of the penalty on different parameters or parameter groups. See MRSP.fit for details.

guessed.active:

Object indexing the covariates that are found to have an effect on the response. If necessary, this refers to groups of coefficients.

guessed.active.coef:

An object with the same structure as slot coef, indexing the atomic coefficients that are nonzero.

guessed.active.groupdiff:

Index of "columnwise" parameter groups that contain at least two different parameter values.

guessed.active.diff:

An index of nonzero pairwise differences between parameters belonging to the same covariate.

df:

Estimated effective degrees of freedom.

tuning:

List of tuning parameter values as supplied to internal function fista.

lambda:

The lambda value, which controls the degree of penalization of most traditional penalties.

lambdaR:

The lambda value used for ridge penalties.

lambdaF:

The lambda value used for fusion penalties.

fusion:

Either FALSE or character string specifying the type of fusion that was used in fusion penalties. Note that those fusion penalties are not yet supported for end-users of MRSP.

gamma:

A numeric that weighs lasso penalties vs CATS lasso penalties. See MRSP for details.

psi:

A numeric that weighs penalties on coefficients of global vs category-specific predictors. See MRSP for details.

eta:

A nobs x K matrix of linear predictor values.

mu:

A nobs x K matrix of estimated probabilities for the response categories.

offset:

Vector of offset values that were added to the linear predictors.

residuals:

Object with residuals. Currently always NULL. Use method residuals instead to compute residuals.

mlfit:

Not to be used by or of interest for end-users.

AIC:

The AIC of the fitted model.

BIC:

The BIC of the fitted model.

Brier:

The Brier score of the fitted model.

threshold:

The numeric threshold used.

refit:

Logical indicating whether the object results from a refitting procedure.

indg:

Not to be used by or of interest for end-users.

indcs:

Not to be used by or of interest for end-users.

model:

The model-object used. To save memory, this is often stored as an expression.

constr:

The identifiability constraint that was used. See MRSP for details.

control:

Object of class "MRSP.control" that contains control information.

fn.val:

The final value of the objective function that was minimized, i.e. the negative penalized loglikelihood: -loglik + lambda*penweights*penalty.

loglik:

Loglikelihood value of the fitted model.

penalty:

Value of the penalty term for the fitted model, weighted with the corresponding tuning parameters and penweights.

iter.count:

Number of iterations until convergence of the proximal gradient algorithm.

best.iter:

Iteration number with the best value of the objective function.

ridgestabil:

Logical indicating whether a small, untuned ridge penalty was applied to all coefficients in order to stabilize otherwise diverging estimates.

name:

A character string specifying the name and thus type of the fitted model.

fisher:

Fisher matrix. Currently not computed and thus always NULL.

arglist:

Not to be used by or of interest for end-users.

call:

The call to MRSP.fit that created this object.

Methods

AIC

signature(object = "MRSP"): Compute the AIC of an MRSP object.

BIC

signature(object = "MRSP"): Compute the BIC of an MRSP object.

coef

signature(object = "MRSP"): See coef-methods.

fitted

signature(object = "MRSP"): ompute fitted values, which for multinomial response correspond to class probabilities P(Y = r). If option convert2hazard=TRUE and a sequential model is used, “discrete” hazard rates P(Y = r | Y >= r) are returned instead.

logLik

signature(object = "MRSP"): Returns the loglikelihood and the (estimated, effective) degrees of freedom.

nobs

signature(object = "MRSP"): Returns the number of invidual observations.

predict

signature(object = "MRSP"): predict(object, newdata, type=c("response", "link"), ...) predicts the response values (type="response") or the linear predictors (type="link") for the observations given in newdata. Additional arguments offset and weights can specify offsets and weights to be used. An argument convert2hazard can be supplied for sequential models, see fitted above.

residuals

signature(object = "MRSP"): Depending on argument type = c("deviance", "pearson"), which is matched via match.arg, deviance or pearson residuals are returned.

show

signature(object = "MRSP"): Print some basic infos about the MRSP object.

summary

signature(object = "MRSP"): Show some slots of an MRSP object which are typically of interest.

Author(s)

Wolfgang Poessnecker

References

Tutz, G., Poessnecker, W., Uhlmann, L. (2015) Variable Selection in General Multinomial Logit Models
Computational Statistics and Data Analysis, Vol. 82, 207-222.
http://www.sciencedirect.com/science/article/pii/S0167947314002709

See Also

MRSP.list

Examples

showClass("MRSP")
## for examples, see ?MRSP

MRSP documentation built on May 29, 2017, 11:36 a.m.