rad_control: Given a policy and (optional) controls, generate a...
In jongbinjung/undi: Test for Unjustified Disparate Impact

Description Usage Arguments Details Value

Given a policy and (optional) controls, generate a rad_control object

1 2	rad_control(pol, fit_fn = c("logit_coef", "gam_coef", "decbin_coef", "logit_avg", "gam_avg"), controls = NULL, use_speedglm = TRUE)

`pol`	a `policy` object
`fit_fn`	string indicating the rad estimation model/procedure used. `_coef` methods use models without interaction between risk and group, and return the coeficient on group membership. `_avg` methods will fit more flexible models (possibly with interactions), and compute average ratios across the population. (TODO: better documentation is expected)
`controls`	character vector of additional controls to consider in the second-stage model
`use_speedglm`	whether or not to use `speedglm`, instead of `stats::glm`, in cases where N > 2P (see details)

speedglm can potentially speed-up computation significantly, but only in cases where the number of rows is somewhat greater than the number of features (specifically, when N > 2P). In terms of FLOPs at each Fisher iteration, stats::glm requires (2np^2 - (2/3)p^3) FLOPS vs, (np^2 + (4/3)p^3) for speedglm.

a rad_control object constructed of

`formula`	the formula used in model fitting
`label`	a character label associated with the model fit type
`grouping`	column name of group, as specified in `pol$grouping`
`fit`	a function of the form f(d, w = NULL, ...) for fitting a model with training data `d`
`pred`	function of the form g(m, d) for generating predictions for data `d` with model `m`
`method`	character string describing the method to use