predict.varbvs {varbvs}

Make predictions from a model fitted by varbvs.

Description

This function predicts outcomes (Y) given the observed variables (X) and observed covariates (Z), and a model fitted using varbvs.

Usage

  ## S3 method for class 'varbvs'
predict(object, X, Z = NULL,
                           type = c("link","response","class"),
                           averaged = TRUE, ...)

Arguments

object

Output of function varbvs.

X

n x p input matrix, in which p is the number of variables, and n is the number of samples for which predictions will be made using the fitted model. X cannot be sparse, and cannot have any missing values (NA).

Z

n x m covariate data matrix, where m is the number of covariates. Do not supply an intercept as a covariate (i.e., a column of ones), because an intercept is automatically included in the regression model. For no covariates, set Z = NULL.

type

Type of prediction to output. The default, "link", gives the linear predictors for family = "binomial", and gives the fitted values for family = "gaussian". For logistic regression (family = "binomial"), there are two alternative predictions: "response" givees the fitted probabilities, and "class" produces the maximum-probability outcome (0 or 1).

averaged

When averaged = TRUE, the predictions are computed by averaging over the hyperparameter settings, treating object$logw as (unnormalized) log-marginal probabilities. (See varbvs for more details about averaging.) When averaged = FALSE, the predictions are returned as a matrix when one row for each data sample, and one column for each hyperparameter setting.

...

Other arguments to generic predict function. These extra arguments are not used here.

Details

Note that the classification probabilities Pr(Y = 1 | X, Z, \theta) are not guaranteed to be calibrated under the variational approximation.

Value

When averaged = TRUE, the output is a vector containing the predicted outcomes for all samples. For family = "binomial", all vector entries are 0 or 1.

When averaged = FALSE, the return value is a matrix with one row for each sample, and one column for each hyperparameter setting.

Author(s)

Peter Carbonetto [e-mail]

References

P. Carbonetto and M. Stephens (2012). Scalable variational inference for Bayesian variable selection in regression, and its accuracy in genetic association studies. Bayesian Analysis 7, 73–108.

See Also

varbvs, summary.varbvs

Examples

  # See help(varbvs) for examples.