ordinis: Fitting Lasso-penalized Using the Coordinate Descent...
In jaredhuling/ordinis: Fits penalized regression models with coordinate descent

Description Usage Arguments Examples

ordinis provides estimation of linear models with the lasso penalty

ordinis(x, y, weights = rep(1, NROW(y)), offset = NULL,
  family = NULL, penalty = c("lasso", "alasso", "mcp", "scad"),
  lambda = numeric(0), alpha = 1, gamma = ifelse(penalty == "scad",
  3.7, 1.4), penalty.factor = NULL, upper.limits = rep(Inf, NCOL(x)),
  lower.limits = rep(-Inf, NCOL(x)), nlambda = 100L,
  lambda.min.ratio = NULL, intercept = TRUE, standardize = TRUE,
  dfmax = nvars, maxit = NULL, tol = NULL, maxit.irls = 25L,
  tol.irls = 0.001)

`x`	The design matrix
`y`	The response vector
`weights`	a vector of weights of length equal to length of `y`. `weights` are NOT standardized or scaled; the user must do this if desired
`offset`	A vector of length `nobs` that is included in the linear predictor
`family`	family of underlying model. Only "gaussian" for continuous responses is available now
`penalty`	a string indicating which penalty to use. `"lasso"`, `"MCP"`, and `"SCAD"` are available
`lambda`	A user provided sequence of λ. If set to `NULL`, the program will calculate its own sequence according to `nlambda` and `lambda_min_ratio`, which starts from λ_0 (with this λ all coefficients will be zero) and ends at `lambda0 * lambda_min_ratio`, containing `nlambda` values equally spaced in the log scale. It is recommended to set this parameter to be `NULL` (the default).
`alpha`	mixing parameter between 0 and 1 for elastic net. `alpha=1` is for the lasso, `alpha=0` is for ridge
`gamma`	parameter for MCP/SCAD. Defaults to the recommended values from the papers corresponding to each penalty
`penalty.factor`	a vector with length equal to the number of columns in x to be multiplied by lambda. by default it is a vector of 1s. `penalty.factor` is NOT scaled
`upper.limits`	a vector of length `ncol(x)` of upper limits for each coefficient. Can be a single value, which will then be applied for each coefficient. Must be non-negative.
`lower.limits`	a vector of length `ncol(x)` of lower limits for each coefficient. Can be a single value, which will then be applied for each coefficient. Cannot be greater than 0.
`nlambda`	Number of values in the λ sequence. Only used when the program calculates its own λ (by setting `lambda = NULL`).
`lambda.min.ratio`	Smallest value in the λ sequence as a fraction of λ_0. See the explanation of the `lambda` argument. This parameter is only used when the program calculates its own λ (by setting `lambda = NULL`). The default value is the same as glmnet: 0.001 if `nrow(x) >= ncol(x)` and 0.05 otherwise.
`intercept`	Whether to fit an intercept in the model. Default is `TRUE`.
`standardize`	Whether to standardize the design matrix before fitting the model. Default is `TRUE`. Fitted coefficients are always returned on the original scale.
`dfmax`	Maximum number of variables allowed in the model
`maxit`	Maximum number of coordinate descent iterations.
`tol`	convergence tolerance parameter.
`maxit.irls`	Maximum number of coordinate descent iterations.
`tol.irls`	convergence tolerance parameter.

set.seed(123)
n = 100
p = 1000
b = c(runif(10, min = 0.1, max = 1), rep(0, p - 10))
x = matrix(rnorm(n * p, sd = 1.5), n, p)
y = drop(x %*% b) + rnorm(n)


## fit lasso model with 100 tuning parameter values
res <- ordinis(x, y)

y2 <- 1 * (y > 0)
y3 <- exp(y)

resb <- ordinis(x, y2, family = "binomial")