policy: Create policy object
In jongbinjung/undi: Test for Unjustified Disparate Impact

Create policy object

policy(formula, data, outcome, train = 0.5, risk = "resp_trt",
  model = names(models()), down_sample = 1, fit1 = NULL, pred1 = NULL,
  fit_ptreat = NULL, pred_ptreat = NULL, ptreat = NULL, resp_ctl = NULL,
  resp_trt = NULL, calibrate = FALSE, save_models = FALSE,
  recast_df = FALSE, seed = round(stats::runif(1) * 10000), ...)

`formula`	a formula in the form of `treatment ~ grouping_variable + other predictors` where the LHS is the treatment column, first element on the RHS is the grouping variable (e.g., Race), and the remainder of the RHS specifies the predictors (features) for treatment/response models
`data`	data frame to use; must include all the columns specified in `formula` and given in the `outcome` parameter
`outcome`	name of outcome column in data
`train`	either (1) a value between 0 and 1 representing the proportion of data to use in training, (2) the name of a column of characters "train" and "test" within `data` to use in splitting the data, or (3) a logical vector of equal length as `nrow(data)` used to index training data
`risk`	One of `"resp_ctl"` or `"resp_trt"`, indicating which treatment regime should be used as the risk score (default: `resp_trt`)
`model`	character of modelling method to use for risk models; run `names(undi::models())` to get list of available models
`down_sample`	(Optional) proportion (between 0 and 1) or number (greater than 1) of rows to sample, if down sampling the data; default is 1 (i.e., use all data)
`fit1`	a function of the form f(formula, data, ...) used for fitting the first-stage model; overides `model`
`pred1`	a function of the form f(model, data, formula) used for generating predictions from the first-stage model; the formula argument can be ignored within the function body, but the function should still accept it; some prediction functions (e.g., glmnet) require the original formula; predictions should be on probability scale, while "risk" will always be on logit scale; overrides `model`
`fit_ptreat`	a function of the form f(formula, data, ...) used for fitting propensity (probability of treatment) models. If not specified, `fit1` is used by default, with the provided `formula` argument.
`pred_ptreat`	a function of the form f(model, data, formula) used for generating propensity predictions. If not specified, `pred1` is used by default.
`ptreat`	(Optional) default value for probability of treatment;
`resp_ctl`	(Optional)
`resp_trt`	(Optional) default value for probability of response = 1 given each treatment regime (`ctl`, `trt`); useful for cases where outcome under certain treatment regimes is deterministic (e.g., probability of finding illegal weapon if NOT frisked is 0)
`calibrate`	whether or not to use platt scaling to calibrate predictions
`save_models`	whether or not fitted models should be returned
`recast_df`	(Optional) whether `data` should be re-cast as a `data.frame`; this might be useful if the modeling function requires a model matrix, and `data` contains character columns — recasting as a data frame will convert all character columns to factors
`seed`	random seed to use
`...`	additional arguments passed to first-stage model fitting function, `fit1` and `fit_ptreat`

policy object

`data`	original data frame, augmented with columns `fold__`, `ptrt__`, `resp_ctl__`, `resp_trt__`, and `risk__`, which contain train/test fold indicators for the first stage, treatment propesinty, first stage model probability predictions under control and treatment, and appropriate risk measure on logit-scale, respectively
`risk_col`	either "resp_ctl" or "resp_trt", indicating which was used for "risk"
`treatment`	name of column from `data` used as treatment indicator
`outcome`	name of column from `data` used as outcome indicator
`grouping`	name of column from `data` used as grouping variable
`features`	additional features used in first stage model
`fit1`	function used to fit first stage model
`pred1`	function used to generate predictions from first stage model
`fit_ptreat`	function used to fit model for treatment propensity
`pred_ptreat`	function used to generate predictions for treatment propensity
`m_*`	if `save_models = TRUE`, each of the fitted models; otherwise set to `NULL`
`call`	the original call