Fit Nuisance Models for Counterfactual Performance Estimation

Fits propensity score and outcome models for use in counterfactual performance estimation.

Usage

fit_nuisance(
  propensity_formula,
  outcome_formula,
  data,
  treatment_level = 0,
  propensity_method = c("glm", "gam"),
  outcome_method = c("glm", "gam"),
  ...
)

Arguments

propensity_formula

Formula for the propensity score model.

outcome_formula

Formula for the outcome model.

data

Data frame containing the variables.

treatment_level

The counterfactual treatment level (default: 0).

propensity_method

Method for propensity score estimation:

"glm": Logistic regression (default)
"gam": Generalized additive model (requires mgcv)

outcome_method

Method for outcome model estimation:

"glm": Logistic regression (default)
"gam": Generalized additive model (requires mgcv)

...

Additional arguments passed to model fitting functions.

Value

An object of class cf_nuisance containing:

propensity: Fitted propensity score model
outcome: Fitted outcome model
treatment_level: Counterfactual treatment level

Details

This function provides a convenient way to fit nuisance models that can then be passed to cf_mse(), cf_auc(), or cf_calibration().

For the outcome model, only observations with the counterfactual treatment level are used for fitting, as required for the conditional loss estimator.

Examples

# Generate example data
set.seed(123)
n <- 500
df <- data.frame(
  x1 = rnorm(n),
  x2 = rnorm(n)
)
df$a <- rbinom(n, 1, plogis(-0.5 + 0.5 * df$x1))
df$y <- rbinom(n, 1, plogis(-1 + df$x1 + 0.5 * df$x2 - 0.5 * df$a))

# Fit nuisance models
nuisance <- fit_nuisance(
  propensity_formula = a ~ x1 + x2,
  outcome_formula = y ~ x1 + x2,
  data = df,
  treatment_level = 0
)

print(nuisance)
#> 
#> Nuisance Models for Counterfactual Performance
#> --------------------------------------------- 
#> Treatment level: 0 
#> 
#> Propensity Score Model:
#>   Method: glm 
#>   Formula: a ~ x1 + x2 
#> 
#> Outcome Model:
#>   Method: glm 
#>   Formula: y ~ x1 + x2 
#>   Fitted on n = 323 observations with A = 0 
#>