gesso.coefnum {gesso}

Get model coefficients with specified number of non-zero interactions

Description

A function to obtain coefficients with target_b_gxe_non_zero specified to control the desired sparsity of interactions in the model.

Usage

gesso.coefnum(cv_model, target_b_gxe_non_zero, less_than = TRUE)

Arguments

cv_model

cross-validated model fit object obtained by using function gesso.cv

target_b_gxe_non_zero

number of non-zero interactions we want to inlcude in the model

less_than

TRUE if we want to control a number of at most non-zero interactions, FALSE if we want to control a number of at least non-zero interactions

Value

A list of model coefficients corresponding to the best model that contains at most or at least target_b_gxe_non_zero non-zero interaction terms.

The target model is selected based on the averaged cross-validation (cv) results: for each pair of parameters lambda=(lambda_1, lambda_2) in the grid and each cv fold we obtain a number of non-zero estimated interaction terms, then average cv results by lambda and choose the tuning parameters corresponding to the minimum average cv loss that have at most or at least target_b_gxe_non_zero non-zero interaction terms. Returned coefficients are obtained by fitting the model on the full data with the selected tuning parameters.

Note that the number of estimated non-zero interactions will only approximately reflect the numbers obtained on cv datasets.

beta_0

estimated intercept value

beta_e

estimated environmental coefficient value

beta_g

a vector of estimated main effect coefficients

beta_gxe

a vector of estimated interaction coefficients

beta_c

a vector of estimated confounders coefficients

Examples

data = data.gen()
model = gesso.cv(data$G_train, data$E_train, data$Y_train)
model_coefficients = gesso.coefnum(model, 5)
gxe_coefficients = model_coefficients$beta_gxe; sum(gxe_coefficients!=0)