predict.grridge: Predictions for new samples

Description Usage Arguments Details Value Author(s) References See Also Examples

Description

Returns predictions for new samples from a grridge object

Usage

1
2
3
## S3 method for class 'grridge'
predict(object, datanew, printpred = FALSE, dataunpennew=NULL,
        responsetest=NULL, recalibrate=FALSE, ...)

Arguments

object

A model object resulted from the grridge function.

datanew

Vector or data frame. Contains the new data. For a data frame: columns are samples, rows are variables (features).

printpred

Boolean. Should the predictions be printed on the screen?

dataunpennew

Vector or data frame. Optional new data for unpenalized variables. NOTE: columns are covariates, rows are samples.

responsetest

Factor, numeric, binary or survival. Response values of test samples. The number of response values should equal ncol(datanew). Only relevant if recalibrate=TRUE.

recalibrate

Boolean. Should the prediction model be recalibrated on the test samples? Only implemented for logistic and linear regression with only penalized covariates.

...

There is no further argument is used.

Details

This function returns predictions of the response using the grridge output. It should be applied to samples NOT used for fitting of the models. About recalibrate: we noticed that recalibration of the linear predictor using a simple regression (with intercept and slope) can improve predictive performance, e.g. in terms of brier score or mean square error (not in terms of AUC, which is rank-based). We recommend to use it for large enough test sets (say >= 25 samples), in particular when one suspects that the test set could have somewhat different properties than the training set. For survival, the proportional hazards are returned, i.e. exp(linear predictor). If survival time predictions are desired, these may be obtained by applying predict to fit objects, which are stored as output of grridge.

Value

A matrix containing the predictions from all models available in grridge.

Author(s)

Mark A. van de Wiel

References

Mark van de Wiel, Tonje Lien, Wina Verlaat, Wessel van Wieringen, Saskia Wilting. (2016). Better prediction by use of co-data: adaptive group-regularized ridge regression. Statistics in Medicine, 35(3), 368-81.

See Also

Cross-validated predictions: grridgeCV. Examples: grridge.

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
#data(dataFarkas)

#firstPartition <- CreatePartition(CpGannFarkas)

#sdsF <- apply(datcenFarkas,1,sd)
#secondPartition <- CreatePartition(sdsF,decreasing=FALSE, uniform=TRUE, grsize=5000)

## Concatenate two partitions
#partitionsFarkas <- list(cpg=firstPartition, sds=secondPartition)

## A list of monotone functions from the corresponding partition
#monotoneFarkas <- c(FALSE,TRUE)

#Hold-out first two samples
#testset <- datcenFarkas[,1:2]
#resptest <- respFarkas[1:2]

#trainingset <- datcenFarkas[,-(1:2)]
#resptraining <- respFarkas[-(1:2)]

#grFarkas <- grridge(trainingset,resptraining,optl=5.680087,
#                    partitionsFarkas,monotone=monotoneFarkas)

## Prediction of the grridge model to the test samples

#Standardize variables in testset, becasuse by default standardizeX = TRUE in grridge function.
#Here, test set is small so we use the summaries of the training set. If the test set
#is large one may opt to standardize wrt to the test set itself. 

#sds <- apply(trainingset,1,sd)
#sds2 <- sapply(sds,function(x) max(x,10^{-5}))
#teststd <- (testset-apply(trainingset,1,mean))/sds2


#yhat <- predict.grridge(grFarkas,teststd)

markvdwiel/GRridge documentation built on May 21, 2019, 12:25 p.m.