glmaag (version 0.0.6)

ss_glmaag: Stability selection for glmaag

Description

Do stability selection for glmaag

Usage

ss_glmaag(y, x, L, nfolds = 5, subn, nsam = 100, beta = 0.15,
  gam = 1, tune = F, est = T, lam1, lam2, w0, adaptl1 = T,
  adaptl2 = T, pind, intercept = T, standardize = T,
  maxiter = 10000, cri = 0.001, fam = "Gaussian", measdev = T,
  type1se = T, parallel = F)

Arguments

y

outcome

x

predictors matrix

L

Laplacian matrix for the first network

nfolds

number of folds used in cross validation to estimate the l1 weights or network tuning, default to be five

subn

number of samples in each subset, default to be n/2 if n<400 and 10sqrt(10) if n>400

nsam

number of subsets, default to be 100

beta

the cut off for instability score

gam

The power of weights of l1 penalty, default to be ones

tune

whether to tune the input network with estimated network or identity matrix, ignored if no input network

est

when there is no input network whether to use estimated network or identiy matrix (elastic net) or mixed the network with estimated network or identity matrix, default to be estimated network

lam1

The tuning parameters for l1 penalty. If not defined, searched by default

lam2

The tuning parameters for quadratic penalty. If not defined, searched by default

w0

Weights for l1 penalty. If not defined, estimated via quadratic penalyzed regression

adaptl1

whether to adapt the l1 penalty, default to be TRUE

adaptl2

whether to adapt the sign for quadratic penalty, default to be TRUE

pind

indicator vector whether to put l1 penalty on the feature, 1 means penalyzed while 0 means not penalyzed, default to be all ones (all penalyzed)

intercept

whether to include intercept. Ignore for Cox regression

standardize

whether to standardize predictors

maxiter

maximum number of iterations, default to be 500

cri

stoppint criterion, default to be 0.001

fam

family for the outcome, can be "Gaussian", "Logistic", and "Cox"

measdev

Whether to use deviance to tune when estimate l1 weight and network sign, default to be deviance. If not, use mean absolue error, area under ROC curve, or concordance index for Gaussian, Logistic, and Cox

type1se

whether to use one standard error or maximum rule for l1 weight estimation and network sign, default to be one standard error rule

parallel

whether to do parallel computing at each subset, need to set up parallel first, default to be FALSE

Value

input

input matrix for predictors

lambda1

searching sequence for l1 penalty parameters

lamdba2

searching sequence for quadratic penalty parameters

lambda1_ss

optimal l1 parameter

lambda2_ss

optimal quadratic parameter

n_ss

number of parameters obtained by the optimal model

ssm

instability score paths

ssf

selection probability paths

intercept_ss

intercept estimated by the optimal model

coef_ss

coefficients estimated by the optimal model

fam

the family of the outcome

References

Meinshausen, N., & B"uhlmann, P. (2010). Stability selection. Journal of the Royal Statistical Society: Series B (Statistical Methodology), 72(4), 417-473.

Liu, H., Roeder, K., & Wasserman, L. (2010). Stability approach to regularization selection (stars) for high dimensional graphical models. In Advances in neural information processing systems (pp. 1432-1440).

Examples

# NOT RUN {
data(sampledata)
data(L0)
y <- sampledata$Y_Gau
x <- sampledata[, 4:6]
mod <- ss_glmaag(y, x, L0[seq_len(3), seq_len(3)], nsam = 3)
# }