sven | R Documentation |

SVEN is an approach to selecting variables with embedded screening using a Bayesian hierarchical model. It is also a variable selection method in the spirit of the stochastic shotgun search algorithm. However, by embedding a unique model based screening and using fast Cholesky updates, SVEN produces a highly scalable algorithm to explore gigantic model spaces and rapidly identify the regions of high posterior probabilities. It outputs the log (unnormalized) posterior probability of a set of best (highest probability) models. For more details, see Li et al. (2020).

```
sven(
X,
y,
w = sqrt(nrow(X))/ncol(X),
lam = nrow(X)/ncol(X)^2,
Ntemp = 3,
Tmax = (log(log(ncol(X))) + log(ncol(X))),
Miter = 50,
wam.threshold = 0.5,
log.eps = -16,
L = 20,
verbose = TRUE
)
```

`X` |
The |

`y` |
The response vector of length |

`w` |
The prior inclusion probability of each variable. Default: |

`lam` |
The slab precision parameter. Default: |

`Ntemp` |
The number of temperatures. Default: 3. |

`Tmax` |
The maximum temperature. Default: |

`Miter` |
The number of iterations per temperature. Default: |

`wam.threshold` |
The threshold probability to select the covariates for WAM. A covariate will be included in WAM if its corresponding marginal inclusion probability is greater than the threshold. Default: 0.5. |

`log.eps` |
The tolerance to choose the number of top models. See detail. Default: -16. |

`L` |
The minimum number of neighboring models screened. Default: 20. |

`verbose` |
If |

SVEN is developed based on a hierarchical Gaussian linear model with priors placed on the regression coefficients as well as on the model space as follows:

`y | X, \beta_0,\beta,\gamma,\sigma^2,w,\lambda \sim N(\beta_01 + X_\gamma\beta_\gamma,\sigma^2I_n)`

`\beta_i|\beta_0,\gamma,\sigma^2,w,\lambda \stackrel{indep.}{\sim} N(0, \gamma_i\sigma^2/\lambda),~i=1,\ldots,p,`

`(\beta_0,\sigma^2)|\gamma,w,p \sim p(\beta_0,\sigma^2) \propto 1/\sigma^2`

`\gamma_i|w,\lambda \stackrel{iid}{\sim} Bernoulli(w)`

where `X_\gamma`

is the `n \times |\gamma|`

submatrix of `X`

consisting of
those columns of `X`

for which `\gamma_i=1`

and similarly, `\beta_\gamma`

is the
`|\gamma|`

subvector of `\beta`

corresponding to `\gamma`

.
Degenerate spike priors on inactive variables and Gaussian slab priors on active
covariates makes the posterior
probability (up to a normalizing constant) of a model `P(\gamma|y)`

available in
explicit form (Li et al., 2020).

The variable selection starts from an empty model and updates the model according to the posterior probability of its neighboring models for some pre-specified number of iterations. In each iteration, the models with small probabilities are screened out in order to quickly identify the regions of high posterior probabilities. A temperature schedule is used to facilitate exploration of models separated by valleys in the posterior probability function, thus mitigate posterior multimodality associated with variable selection models. The default maximum temperature is guided by the asymptotic posterior model selection consistency results in Li et al. (2020).

SVEN provides the maximum a posteriori (MAP) model as well as the weighted average model
(WAM). WAM is obtained in the following way: (1) keep the best (highest probability) `K`

distinct models `\gamma^{(1)},\ldots,\gamma^{(K)}`

with

`\log P\left(\gamma^{(1)}|y\right) \ge \cdots \ge \log P\left(\gamma^{(K)}|y\right)`

where `K`

is chosen so that
`\log \left\{P\left(\gamma^{(K)}|y\right)/P\left(\gamma^{(1)}|y\right)\right\} > \code{log.eps}`

;
(2) assign the weights

`w_i = P(\gamma^{(i)}|y)/\sum_{k=1}^K P(\gamma^{(k)}|y)`

to the model `\gamma^{(i)}`

; (3) define the approximate marginal inclusion probabilities
for the `j`

th variable as

`\hat\pi_j = \sum_{k=1}^K w_k I(\gamma^{(k)}_j = 1).`

Then, the WAM is defined as the model containing variables `j`

with
`\hat\pi_j > \code{wam.threshold}`

. SVEN also provides all the top `K`

models which
are stored in an `p \times K`

sparse matrix, along with their corresponding log (unnormalized)
posterior probabilities.

A list with components

`model.map` |
A vector of indices corresponding to the selected variables in the MAP model. |

`model.wam` |
A vector of indices corresponding to the selected variables in the WAM. |

`model.top` |
A sparse matrix storing the top models. |

`beta.map` |
The ridge estimator of regression coefficients in the MAP model. |

`beta.wam` |
The ridge estimator of regression coefficients in the WAM. |

`mip.map` |
The marginal inclusion probabilities of the variables in the MAP model. |

`mip.wam` |
The marginal inclusion probabilities of the variables in the WAM. |

`pprob.map` |
The log (unnormalized) posterior probability corresponding to the MAP model. |

`pprob.top` |
A vector of the log (unnormalized) posterior probabilities corresponding to the top models. |

`stats` |
Additional statistics. |

Dongjin Li and Somak Dutta

Maintainer:
Dongjin Li <dongjl@iastate.edu>

Li, D., Dutta, S., Roy, V.(2020) Model Based Screening Embedded Bayesian Variable Selection for Ultra-high Dimensional Settings http://arxiv.org/abs/2006.07561

```
n=50; p=100; nonzero = 3
trueidx <- 1:3
nonzero.value <- 5
TrueBeta <- numeric(p)
TrueBeta[trueidx] <- nonzero.value
rho <- 0.6
x1 <- matrix(rnorm(n*p), n, p)
X <- sqrt(1-rho)*x1 + sqrt(rho)*rnorm(n)
y <- 0.5 + X %*% TrueBeta + rnorm(n)
res <- sven(X=X, y=y)
res$model.map # the MAP model
res$model.wam # the WAM
res$mip.map # the marginal inclusion probabilities of the variables in the MAP model
res$mip.wam # the marginal inclusion probabilities of the variables in the WAM
res$pprob.top # the log (unnormalized) posterior probabilities corresponding to the top models.
res$beta.map # the ridge estimator of regression coefficients in the MAP model
res$beta.wam # the ridge estimator of regression coefficients in the WAM
```

Embedding an R snippet on your website

Add the following code to your website.

For more information on customizing the embed code, read Embedding Snippets.