Multi-Step Adaptive MCP-Net

## Usage

```
msamnet(
x,
y,
family = c("gaussian", "binomial", "poisson", "cox"),
init = c("mnet", "ridge"),
gammas = 3,
alphas = seq(0.05, 0.95, 0.05),
tune = c("cv", "ebic", "bic", "aic"),
nfolds = 5L,
ebic.gamma = 1,
nsteps = 2L,
tune.nsteps = c("max", "ebic", "bic", "aic"),
ebic.gamma.nsteps = 1,
scale = 1,
eps = 1e-04,
max.iter = 10000L,
penalty.factor.init = rep(1, ncol(x)),
seed = 1001,
parallel = FALSE,
verbose = FALSE
)
```

## Arguments

- x
Data matrix.

- y
Response vector if

`family`

is`"gaussian"`

,`"binomial"`

, or`"poisson"`

. If`family`

is`"cox"`

, a response matrix created by`Surv`

.- family
Model family, can be

`"gaussian"`

,`"binomial"`

,`"poisson"`

, or`"cox"`

.- init
Type of the penalty used in the initial estimation step. Can be

`"mnet"`

or`"ridge"`

.- gammas
Vector of candidate

`gamma`

s (the concavity parameter) to use in MCP-Net. Default is`3`

.- alphas
Vector of candidate

`alpha`

s to use in MCP-Net.- tune
Parameter tuning method for each estimation step. Possible options are

`"cv"`

,`"ebic"`

,`"bic"`

, and`"aic"`

. Default is`"cv"`

.- nfolds
Fold numbers of cross-validation when

`tune = "cv"`

.- ebic.gamma
Parameter for Extended BIC penalizing size of the model space when

`tune = "ebic"`

, default is`1`

. For details, see Chen and Chen (2008).- nsteps
Maximum number of adaptive estimation steps. At least

`2`

, assuming adaptive MCP-net has only one adaptive estimation step.- tune.nsteps
Optimal step number selection method (aggregate the optimal model from the each step and compare). Options include

`"max"`

(select the final-step model directly), or compare these models using`"ebic"`

,`"bic"`

, or`"aic"`

. Default is`"max"`

.- ebic.gamma.nsteps
Parameter for Extended BIC penalizing size of the model space when

`tune.nsteps = "ebic"`

, default is`1`

.- scale
Scaling factor for adaptive weights:

`weights = coefficients^(-scale)`

.- eps
Convergence threshhold to use in MCP-net.

- max.iter
Maximum number of iterations to use in MCP-net.

- penalty.factor.init
The multiplicative factor for the penalty applied to each coefficient in the initial estimation step. This is useful for incorporating prior information about variable weights, for example, emphasizing specific clinical variables. To make certain variables more likely to be selected, assign a smaller value. Default is

`rep(1, ncol(x))`

.- seed
Random seed for cross-validation fold division.

- parallel
Logical. Enable parallel parameter tuning or not, default is FALSE. To enable parallel tuning, load the

`doParallel`

package and run`registerDoParallel()`

with the number of CPU cores before calling this function.- verbose
Should we print out the estimation progress?

## Author

Nan Xiao <https://nanx.me>

## Examples

```
dat <- msaenet.sim.gaussian(
n = 150, p = 500, rho = 0.6,
coef = rep(1, 5), snr = 2, p.train = 0.7,
seed = 1001
)
msamnet.fit <- msamnet(
dat$x.tr, dat$y.tr,
alphas = seq(0.3, 0.9, 0.3),
nsteps = 3L, seed = 1003
)
#> Warning: ncvreg() is intended for pathwise optimization, not for single values of lambda.
#> 1. You are strongly encouraged to fit a path and extract the solution at the lambda value of interest, rather than use ncvreg() in this way.
#> 2. In particular, if you are using the MCP or SCAD penalties, be aware that you greatly increase your risk of converging to an inferior local maximum if you do not fit an entire path.
#> 3. You may wish to look at the ncvfit() function, which is intended for non-path (i.e., single-lambda) optimization and allows the user to supply initial values.
#> Warning: ncvreg() is intended for pathwise optimization, not for single values of lambda.
#> 1. You are strongly encouraged to fit a path and extract the solution at the lambda value of interest, rather than use ncvreg() in this way.
#> 2. In particular, if you are using the MCP or SCAD penalties, be aware that you greatly increase your risk of converging to an inferior local maximum if you do not fit an entire path.
#> 3. You may wish to look at the ncvfit() function, which is intended for non-path (i.e., single-lambda) optimization and allows the user to supply initial values.
#> Warning: ncvreg() is intended for pathwise optimization, not for single values of lambda.
#> 1. You are strongly encouraged to fit a path and extract the solution at the lambda value of interest, rather than use ncvreg() in this way.
#> 2. In particular, if you are using the MCP or SCAD penalties, be aware that you greatly increase your risk of converging to an inferior local maximum if you do not fit an entire path.
#> 3. You may wish to look at the ncvfit() function, which is intended for non-path (i.e., single-lambda) optimization and allows the user to supply initial values.
#> Warning: ncvreg() is intended for pathwise optimization, not for single values of lambda.
#> 1. You are strongly encouraged to fit a path and extract the solution at the lambda value of interest, rather than use ncvreg() in this way.
#> 2. In particular, if you are using the MCP or SCAD penalties, be aware that you greatly increase your risk of converging to an inferior local maximum if you do not fit an entire path.
#> 3. You may wish to look at the ncvfit() function, which is intended for non-path (i.e., single-lambda) optimization and allows the user to supply initial values.
print(msamnet.fit)
#> Call: msamnet(x = dat$x.tr, y = dat$y.tr, alphas = seq(0.3, 0.9, 0.3),
#> nsteps = 3L, seed = 1003)
#> Df Lambda Gamma Alpha
#> 1 3 0.4095024 3 0.9
msaenet.nzv(msamnet.fit)
#> [1] 2 4 35
msaenet.fp(msamnet.fit, 1:5)
#> [1] 1
msaenet.tp(msamnet.fit, 1:5)
#> [1] 2
msamnet.pred <- predict(msamnet.fit, dat$x.te)
msaenet.rmse(dat$y.te, msamnet.pred)
#> [1] 2.909146
plot(msamnet.fit)
```