mcr_set: Compute MR & MCR

In aaronjfisher/mcr: Computes model class reliance, i.e., the most and least that models from a class may rely on variables of interest

Description

Main functions for computing model reliance (MR) and model class reliance (MCR). For further details, see getMCR_internal. The function precompute_mcr_objects_and_functions computes objects that are used within the MCR search procedure.

Usage

get_MR_general(model, precomputed = NA, ...)

precompute_mcr_objects_and_functions(X, y, p1, model_class_loss = NULL,
  minimize_weighted_loss = NULL, get_loss = NULL, nrep_sample = 2,
  loop_ind_args = NULL)

get_empirical_MCR(eps, precomputed = NA, ...)

Arguments

model	model for which MR should be evaluated. If applicable, this should match the output of minimize_weighted_loss, and be an input for get_loss.
precomputed	output from precompute_mcr_objects_and_functions
...	passed to precompute_mcr_objects_and_functions (and getMCR_internal for advanced options).
X	a matrix or dataframe of covariates
y	outcome vector
p1	numeric index marking which columns of X to compute importance for.
model_class_loss	(optional) Signify a preset model class and loss function. This package currently supports 'linear_mse' for linear regression with the squared error loss; 'kernel_mse' for regression in a reproducing kernel Hilbert space, with the squared error loss; and 'linear_hinge' for linear classification with the hinge loss (y = -1 or 1). If this option is not set, both minimize_weighted_loss and get_loss must be set.
minimize_weighted_loss	(optional) a function with named inputs X, y, case.weights, and (optionally) start. The function should export a model object which minimizes the sum of losses for the dataset (y,X). Any model format is acceptable, as long as it can be read by get_loss. The (optional) input start should also take the same format as the output of minimize_weighted_loss. Regularization constraints should also be accounted for in this function.
get_loss	(optional) a function which takes named inputs model, y, and X, and returns a vector of losses. The input model should be in the same format of the returned value of minimize_weighted_loss.
nrep_sample	an integer between 2 and length(y) determining the level of approximation for empirical MR. If nrep_sample=2, e_divide is used. If nrep_sample=length(y), all combinations are computed. See get_full_sample.
loop_ind_args	A list of arguments not changing over the binary search (e.g., crossprod(X)).
eps	performance threshold for MCR, on an absolute scale.

Value

precompute_mcr_objects_and_functions returns a list containing

• suff_stats - a list of precomuted objects that do not change over iterations of the MCR binary search loop.

• get_ld_h_min - (advanced or internal use) a function that minimizes of equations in the form of h_γ, which has loop-dependent ("ld") arguments. Takes arguments loop_dep_args (optional), suff_stats (above), s (-1 for MCR-, +1 for MCR+), and gam (coefficient in h_γ, see details in main paper). If only computing MR (see get_MR_general), this element may be NULL.

• get_MR - a function taking model and suff_stats as input (see above), and which computes the MR of the model using the precomputed object suff_stats.

get_empirical_MCR returns a list containing the MCR range, the epsilon value, and more detailed results (minus and plus) for each end of the MCR interval (see getMCR_internal for more details).

aaronjfisher/mcr documentation built on Jan. 2, 2020, 4:38 p.m.