iregnet: Fit interval censored AFT models with elastic net...
In anujkhare/iregnet: Regularized interval regression

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

Fit accelerated failure time models using interval censored data via elastic net penalized maximum likeihood. Solutions are computed using coordinate descent for a path of values of the regularization parameter lambda. Supports gaussian, logistic and extreme value distributions.

iregnet(x, y, family = c("gaussian", "logistic", "loggaussian",
  "loglogistic", "extreme_value", "exponential", "weibull"), alpha = 1,
  lambda = NULL, num_lambda = 100, intercept = TRUE,
  standardize = TRUE, scale_init = NA, estimate_scale = TRUE,
  maxiter = 1000, threshold = 0.001, unreg_sol = TRUE,
  eps_lambda = NA, debug = 0)

`x`	Input matrix of covariates with dimension n_obs * n_vars, with nvars ≥ 2. Sparse matrices are not supported.
`y`	Response variable. It can take two forms: 2 column real matrix with NAs denoting a censored observation `Surv` object. Supported `types` of `Surv` are 'left', 'right', 'interval' and 'interval2'.
`family`	The distribution to fit. It can be one of "gaussian", "logistic", "loggaussian", "loglogistic", "extreme_value", "exponential". Partial matching is allowed. Default: "gaussian"
`alpha`	Elastic net mixing parameter, with 0 ≤ α ≤ 1. The elastic net penalty is defined as in `glmnet`: 0.5 (1-α) \\|β\\|_2^2 \| + α \\|β\\|_1* alpha=1 is the lasso penalty, and alpha=0 is ridge penalty. Default: 1
`lambda`	Vector containing the path of decreasing regularization parameter lambda values. If not supplied, the function will calculate a lambda path of length `num_lambda` itself. NOTE: The lambda values are scaled because of the nuisance parameter, and hence not directly comparable to those of other packages like `glmnet`. Default: `NA`
`num_lambda`	The number of lambda values calculated by the function. Ignored if `lambda` is supplied by the user. Default: 100
`intercept`	`TRUE` if an intercept is to be fit, otherwise `FALSE`. Intercept is calculated by appending a column of 1s to `x`. Default: `TRUE`
`standardize`	`TRUE` if `x` must be standardized before fit, otherwise `FALSE`. Calculated `beta` are scaled to original scale before returning. Default: `TRUE`
`scale_init`	Initial value of the scale parameter to use. If not supplied, a suitable value is calculated depending on the distribution. Default: NA
`estimate_scale`	`TRUE` if `scale` is to be estimated. To use a fixed value `scale0` for `scale`, set `scale_init=scale0 , estimate_scale=FALSE`. See examples. Default: `TRUE`
`maxiter`	Maximum number of iterations over data per lambda value. Default: 1e3
`threshold`	The convergence threshold for coordinate descent. The inner loop continues until the absolute update in each parameter is greater than `threshold`. Default: 1e-4
`unreg_sol`	`TRUE` if the final solution computed must be unregularized. Overwritten to `FALSE` if n_vars > n_obs. Only used if `lambda_path` is not specified. Default: `TRUE`
`eps_lambda`	The ratio of the minimum value of `lambda` to the (calculated) maximum value, in case no lambda is supplied. `num_lambda` `lambda` values are calculated between `lambda_max` and `lambda_min` on the log scale. Default: 0.0001 if n_vars < n_obs, 0.1 otherwise.
`debug`	`TRUE` if code debugging messages must be printed. Default: `FALSE`

At each regularization parater value lambda, cyclic coordinate descent is used to update the parameters until convergence. The intercept and the scale parameter are never regularized. The obtained solution is used to initialize the parameters at the next lambda value.

Returns a S3 object iregnet with the following elements:

`beta`	Matrix of size `(n_vars+1) * num_lambda` containing intercept, coefficients of `X` for each `lambda` in the fit model.
`call`	Copy of the call that produced this object.
`lambda`	Vector of size `num_lambda` of (calculated or supplied) regularization parameter `lambda` values.
`loglik`	Vector of size `num_lambda` of log-likelihoods of the fit at each `lambda` value, excluding the contribution of the penalty terms.
`num_lambda`	Number of `lambda` values.
`n_iters`	Vector of size `num_lambda` of number of iterations taken at each `lambda`.
`scale`	Vector of size `num_lambda` of estimated scale at each `lambda` value, if `estimate_scale == TRUE`. Same as `scale_init` otherwise.
`scale_init`	Initial value (calculated or supplied) of `scale`.
`estimate_scale`	`TRUE` if the `scale` was estimated.
`error_status`	The error status. `0` denotes no errors. `-1` denotes that convergence was not reached in `maxiter`.

Friedman, J., Hastie, T. and Tibshirani, R. (2008) Regularization Paths for Generalized Linear Models via Coordinate Descent, http://www.stanford.edu/~hastie/Papers/glmnet.pdf

Simon, N., Friedman, J., Hastie, T., Tibshirani, R. (2011) Regularization Paths for Cox's Proportional Hazards Model via Coordinate Descent, Journal of Statistical Software, Vol. 39(5) 1-13 http://www.jstatsoft.org/v39/i05/

Anuj Khare, Toby Dylan Hocking, Jelle Goeman.
Maintainer: Anuj Khare khareanuj18@gmail.com

predict.iregnet, cv.iregnet, plot.iregnet

# y can be a 2 column matrix.
set.seed(10)
X <- matrix(rnorm(50), 10, 5)
y <- matrix(rnorm(20), 10, 2)
y <- t(apply(y, 1, sort)) # intervals must be non-decreasing
fit1 <- iregnet(X, y)

# Surv objects from survival are also supported.
data("ovarian", package="survival")
library(survival)
X <- cbind(ovarian$ecog.ps, ovarian$rx)
y <- Surv(ovarian$futime, ovarian$fustat)
fit2 <- iregnet(X, y)

# Log-Gaussian is same as Gaussian with log-transformed data
X <- cbind(ovarian$ecog.ps, ovarian$rx)
y <- Surv(ovarian$futime, ovarian$fustat)
y_log <- Surv(log(ovarian$futime), ovarian$fustat)
fit3 <- iregnet(X, y_log, "gaussian", threshold=1e-2)
fit4 <- iregnet(X, y, "loggaussian", threshold=1e-2)

# Scale parameter can be fixed by setting the estimate_scale flag.
set.seed(10)
X <- matrix(rnorm(50), 10, 5)
y <- matrix(rnorm(20), 10, 2)
y <- t(apply(y, 1, sort)) # intervals must be non-decreasing
fit5 <- iregnet(X, y, scale_init=1, estimate_scale=FALSE)