IPW Estimator of the Causal Effects of Additive Modified Treatment Policies

IPW Estimator of the Causal Effects of Additive Modified Treatment Policies

ipw_shift(
  W,
  A,
  Y,
  delta = 0,
  n_bins = make_bins(A, "hist"),
  cv_folds = 10L,
  lambda_seq,
  ...,
  bin_type = c("equal_range", "equal_mass"),
  selector_type = c("dcar", "plateau", "gcv", "all")
)

Arguments

W

A matrix, data.frame, or similar containing a set of baseline covariates.

A

A numeric vector corresponding to a exposure variable. The parameter of interest is defined as a location shift of this quantity.

Y

A numeric vector of the observed outcomes.

delta

A numeric value indicating the shift in the exposure to be used in defining the target parameter. This is defined with respect to the scale of the exposure (A).

n_bins

A numeric, scalar or vector, indicating the number of bins into which the support of A is to be partitioned for constructing conditional density estimates.

cv_folds

A numeric giving the number of folds to be used for cross-validation. Note that this form of sample splitting is used for the selection of tuning parameters by empirical risk minimization, not for the estimation of nuisance parameters (i.e., to relax regularity conditions).

lambda_seq

A numeric sequence of the regularization parameter (L1 norm of HAL coefficients) to be used in fitting HAL models.

...

Additional arguments for model fitting to be passed directly to haldensify.

bin_type

A character indicating the strategy to be used in creating bins along the observed support of A. For bins of equal range, use "equal_range"; to ensure each bin has the same number of observations, use instead "equal_mass". For more information, see documentation of grid_type in haldensify.

selector_type

A character indicating the selection strategy for identifying an efficent IPW estimator. The choices include "gcv" for global cross-validation, "dcar" for solving the EIF equation, and "plateau" for agnostic approaches (1) balancing changes in the IPW estimate and its standard error (adapting Lepski's method) and (2) a plateau detector for inflection points in the IPW estimator's trajectory. The option "all" runs all three selection strategies while sharing redundant computation between each.

Examples

# simulate data
n_obs <- 50
W1 <- rbinom(n_obs, 1, 0.6)
W2 <- rbinom(n_obs, 1, 0.2)
W3 <- rpois(n_obs, 3)
A <- rpois(n_obs, 3 * W1 - W2 + 2 * W1 * W2 + 4)
Y <- rbinom(n_obs, 1, plogis(A + W1 + W2 - W3 - W1 * W3))

# fit the IPW estimator
est_ipw <- ipw_shift(
  W = cbind(W1, W2, W3), A = A, Y = Y,
  delta = 0.5, cv_folds = 2L,
  n_bins = 5L, bin_type = "equal_range",
  lambda_seq = exp(seq(-1, -10, length = 100L)),
  # arguments passed to hal9001::fit_hal()
  max_degree = 3,
  smoothness_orders = 0,
  num_knots = NULL,
  reduce_basis = 1 / sqrt(n_obs)
)
#> Warning: Some fit_control arguments are neither default nor glmnet/cv.glmnet arguments: n_folds; 
#> They will be removed from fit_control
#> 4% of observations outside training support...predictions trimmed.
#> Warning: Some fit_control arguments are neither default nor glmnet/cv.glmnet arguments: n_folds; 
#> They will be removed from fit_control
#> Warning: Dropping reduce_basis; only applies if smoothness_orders = 0