2. Data cleaning

All supplied columns must be numeric and contain no missing values. Otherwise, training or prediction stops with an error identifying the affected columns. This also applies to unused columns, so remove them before passing your data to mcf. The check runs before the cleaning step, regardless of the setting of dc_clean_data.

For discrete treatments, mcf internally converts the treatment categories to consecutive integers starting at zero. If treatment information is required for prediction, use the same treatment labels as in the training data. For policy evaluation, treatment codes must instead correspond to the order of the policy scores specified in var_polscore_name.

Unordered features require at least three distinct training values. Put binary features in var_x_name_ord. Heterogeneity variables in var_z_name_* are added to the corresponding forest-feature lists automatically.

2.1. Parameter overview

Parameter	Description
dc_clean_data	Drops missing observations and unused variables at the cleaning stage; prior validation still applies. Default: True.
dc_screen_covariates	Screens features, including those without variation. Default: True.
dc_check_perfectcorr	Screens near-duplicates using absolute Pearson correlation above 0.999. Protected variables may remain. Default: True.
dc_min_dummy_obs	Minimum count for either value of a binary feature, not only literal 0/1. Use 1 to disable this filter; 0 restores the default of 10.

See the API for details.

2.2. Example

The following block illustrates a configuration; call train() and predict() as in Getting started to obtain estimates.

from mcf.example_data import example_data
from mcf.mcf_main import ModifiedCausalForest

# Generate example data using the built-in function `example_data()`
training_df, prediction_df, name_dict = example_data()

my_mcf = ModifiedCausalForest(
    var_y_name="outcome",
    var_d_name="treat",
    var_x_name_ord=["x_cont0", "x_cont1", "x_ord1"],
    # Parameters for data cleaning:
    dc_clean_data=True,
    dc_screen_covariates=True,
    dc_check_perfectcorr=False,
    dc_min_dummy_obs=100
)