ECLasso.fit: Use ADMM to find solution pairs and approximate parameters in...

In umich-biostatistics/ECLasso: Solving Equality-Constrained Generalized Linear Models or Cox Regression with Lasso Penalty via Alternating Direction Method of Multipliers (ADMM)

Description

Use ADMM to find solution pairs and approximate parameters in Gaussian, Binomial, Poisson, and Cox models.

Usage

ECLasso.fit(x, y, lambda_seq = NULL, family = c("gaussian", "binomial",
  "poisson", "cox"), rho = 1, intercept = TRUE, equality = FALSE,
  censor = NULL, C = NULL, d = NULL, inexact = FALSE,
  tol = 1e-04, maxit = 1000, nlambda = 100,
  lambda.min.ratio = ifelse(nobs < nvars, 0.01, 1e-04), min.abs = 0,
  lower.limits = -Inf, upper.limits = Inf, penalty.factor = 1,
  toplot = TRUE, ParamCutoff = NULL, exclude = rep(FALSE, nvars))

Arguments

x	The design matrix of the data.
y	The response vector; must contain the same amount of values as rows of x.
lambda_seq	User-specified sequence of lambdas the algorithm forced to iterate over, default NULL.
family	The type of model the data comes from; gaussian, binomial, poisson, or Cox.
rho	Weight given to equality constraint vs. loss function, default is 1 aka equal weighting.
intercept	Determines whether to add an intercept to the design matrix.
equality	Forces the parameters to sum to zero if TRUE.
censor	A 0/1 vector for Cox family only, where 1 is exact and 0 is right-censored; default is NULL.
C	Part of equality constraints, set to 1 if equality is TRUE, 0 otherwise.
d	Part of equality constraints, always set to 0 unless user-defined.
inexact	If true, solves a quadratic approximation of the objective function at each iteration; required to be TRUE for "cox" family.
tol	Determines stopping criteria for the algorithm.
maxit	Half the max amount of time each iteration allowed to run.
nlambda	The number of lambdas the function iterates over, should be a multiple of 50.
lambda.min.ratio	Defines what fraction of the maximum lambda the minimum is.
min.abs	Any covariate effect with smaller absolute value than this is set to zero during iteration.
lower.limits	Lower boundary of parameter, any iteration with term(s) below this is reset to this value.
upper.limits	Upper boundary of parameter, any iteration with term(s) above this is reset to this value.
penalty.factor	Penalty factor for shrinkage function, involved in z update.
toplot	Determines if a plot of the solution path should be outputted
ParamCutoff	Specify a cutoff above which predictors are returned at completion, if NULL by default, returns the 6 predictors with greatest magnitude

Value

Solution: parameter estimates at each iteration, time: total elapsed program time, Iter: Time taken for each iteration (consider increasing maxit if this is at maximum), Lambda: the lambda for each iteration, Intercept: intercept value if specified TRUE.

Examples

c0=0.1
beta1=1
beta2=-1
beta3=0.5
beta4=-0.5
betazeroes<-rep(0,6)
betavector<-c(beta1,beta2,beta3,beta4,t(betazeroes))
n=200
xdata<-matrix(rnorm(n*length(betavector)),ncol=length(betavector))
times=rexp(n,c0*exp(rowSums(t(t(xdata)*betavector))))
time.censor=rexp(n,c0*exp(beta1*xdata[,1]))
censorv=ifelse(times<time.censor, 1, 0)
time <- ifelse(times<time.censor, times, time.censor)
ECLasso.fit(x=xdata,y=time,family="cox",intercept=FALSE,equality=TRUE,censor=censorv,inexact=TRUE)

umich-biostatistics/ECLasso documentation built on Jan. 29, 2020, 9:12 p.m.