varbvs.properties: Accessing Properties of Fitted varbvs Models

In varbvs: Large-Scale Bayesian Variable Selection Using Variational Methods

Accessing Properties of Fitted varbvs Models

Description

All these functions are methods for class "varbvs" objects.

Usage

## S3 method for class 'varbvs'
nobs(object, ...)
## S3 method for class 'varbvs'
case.names(object, ...)
## S3 method for class 'varbvs'
variable.names(object, full = FALSE,
                                include.threshold = 0.01, ...)
## S3 method for class 'varbvs'
labels(object, ...)
## S3 method for class 'varbvs'
coef(object, ...)
## S3 method for class 'varbvsmix'
coef(object, ...)
## S3 method for class 'varbvs'
confint(object, parm, level = 0.95, ...)
## S3 method for class 'varbvs'
fitted(object, ...)
## S3 method for class 'varbvs'
resid(object, type = c("deviance","response"), ...)
## S3 method for class 'varbvs'
residuals(object, type = c("deviance","response"), ...)
## S3 method for class 'varbvs'
deviance(object, ...)

Arguments

object	An object inheriting from class varbvs, usually the result of calling function varbvs.
full	logical; if TRUE, names of all variables (columns of X) are returned, including variables that have zero probability of being included in the regression model.
include.threshold	When full = FALSE, names of all variables (columns of X) with "averaged" posterior inclusion probability greater than include.threshold are returned.
parm	Confidence intervals are computed for these selected variables. These may either be specified as numbers (column indices of varbvs input matrix X) or names (column names of X). If not specified, confidence intervals will be computed for the top 5 variables by posterior inclusion probability. Confidence intervals are not provided for covariates (columns of Z); see below for details.
level	Size of confidence level.
type	Type of residuals to be returned. This argument is only relevant for logistic regression models (family = "binomial"). See varbvs for more details about the two available types of residuals for logistic regression.
...	Further arguments passed to or from other methods.

Details

The generic accessor functions nobs, case.names, variable.names and labels can be used to extract various useful properties of the fitted varbvs model. Method labels, in particular, returns the names of the candidate variables (columns of X) which may be used, for example, to plot posterior inclusion probabilities or effect estimates.

coef returns a matrix containing the posterior estimates of the regression coefficients at each hyperparameter setting, as well as an additional column containing "averaged" coefficient estimates.

confint returns confidence intervals (also, equivalently in this case, "credible intervals") for all selected variables parm. These are conditional confidence intervals; that is, conditioned on each variable being included in the regression model.

The confint return value is different from the usual confidence interval (e.g., for an lm result) because a confidence interval is provided for each hyperparameter setting, as well as an additional "averaged" confidence interval. The confidence intervals are returned a list, with one list element per selected variable, and each list element is a matrix with columns giving lower and upper confidence limits for each hyperparameter setting, as well as the averaged limits.

Note that confidence intervals cannot currently be requested for covariates (columns of Z).

fitted returns a matrix containing the predicted (or "fitted") values of the outcome at each hyperparameter setting. For the logistic regression model (family = "binomial"), each matrix entry gives the probability that the binary outcome is equal to 1.

Likewise, resid and residuals each return a matrix containing the model residuals at each hyperparameter setting.

deviance returns the deviance for the fitted model at each hyperparameter setting.

See Also

varbvs, nobs, case.names, variable.names, labels, coef, coef, fitted, residuals, deviance

varbvs documentation built on June 7, 2023, 5:43 p.m.