kFoldCV: kFoldCV

In jhorzek/laremm: LASSO Regularization in mxModels

kFoldCV

Description

Note: laremm is based on the R package regsem. Because of the early status of laremm, it is recommended to use regsem instead! kFoldCV uses k fold cross-validation with fitRegModels

Usage

kFoldCV(k = 5, model, model_type = "mxModel", fitfun = "FIML",
  data_type = "raw", pen_type = "lasso", pen_on = "none",
  selectedDrifts = "none", driftexpo = TRUE, selectedA = "none",
  selectedS = "none", pen_start = 0, pen_end = 1,
  pen_stepsize = 0.01, zeroThresh = 0.001, setZero = FALSE)

Arguments

k	specifies the number of splits (e.g. k = 5 for 5-fold-CV)
model	mxModel with the full data set in model$data. The data set will be split by kFoldCV
model_type	specify the type of model provided: only mxModel supported
fitfun	fitfunction to be used in the fitting procedure. Currently only FIML implemented
data_type	type of data in the model. Only "raw" supported
pen_on	string vector with matrices that should be regularized. Possible are combinations of "A", "S", "DRIFT"
selectedDrifts	drift values to regularize. Possible are "all", "cross", "auto" or providing a matrix of the same size as the drift matrix with ones for every parameter to regularize and 0 for every non-regularized parameter
driftexpo	specifiy if the regularization will be performed on the raw drift matrix or on the exponential of the drift matrix (discrete time parameters)
selectedA	A values to regularize. Possible are "all", or providing a matrix of the same size as the A matrix with ones for every parameter to regularize and 0 for every non-regularized parameter
selectedS	S values to regularize. Possible are "all", or providing a matrix of the same size as the S matrix with ones for every parameter to regularize and 0 for every non-regularized parameter
pen_start	lowest penalty value to evaluate. Recommended: 0
pen_end	highest penalty value to evaluate
pen_stepsize	increse of penalty with each iteration. e.g. if pen_start = 0, pen_end = 1, pen_stepsize = .1, fitRegModels will iterate over pen = 0, pen = .1, pen = .2, ...
zeroThresh	threshold for evaluating regularized parameters as zero. Default is .001 similar to regsem
setZero	should parameters below zeroThresh be set to zero in all fit calculations. Default is FALSE, similar to regsem
penalty_type	so far only "lasso" implemented
DRIFT_dt	provide the discrete time points for which the drift will be regularized. A vector with multiple values is possible

Author(s)

Jannik Orzek

Examples

# The following example is taken from the regsem help to demonstrate the equivalence of both methods:

library(lavaan)
library(OpenMx)
# put variables on same scale for regsem
HS <- data.frame(scale(HolzingerSwineford1939[,7:15]))

# define variables:
latent = c("f1")
manifest = c("x1","x2","x3","x4","x5", "x6", "x7", "x8", "x9")

# define paths:
loadings <- mxPath(from = latent, to = manifest, free = c(F,T,T,T,T,T,T,T,T), values = 1)
lcov <- mxPath(from = latent, arrows = 2, free = T, values = 1)
lmanif <- mxPath(from = manifest, arrows =2 , free =T, values = 1)

# define model:
myModel <- mxModel(name = "myModel", latentVars = latent, manifestVars = manifest, type = "RAM",
                   mxData(observed = HS, type = "raw"), loadings, lcov, lmanif,
                   mxPath(from = "one", to = manifest, free = T)
)

fit_myModel <- mxRun(myModel)
summary(fit_myModel)

# create regularized model:

selectedA <- matrix(0, ncol = ncol(fit_myModel$A$values), nrow = nrow(fit_myModel$A$values))
selectedA[c(2,3,7,8,9),10] <-1


kFolRreg_model <- kFoldCV(model = fit_myModel, model_type = "mxModel", fitfun = "FIML",
                          pen_on = "A", selectedA = selectedA,
                          pen_start = 0, pen_end = .05, pen_stepsize = .01,
                          k = 5
                          )
summary(kFolRreg_model)
# inspect results:
kFolRreg_model$`CV results`

jhorzek/laremm documentation built on Sept. 16, 2022, 12:06 p.m.