EffectXshift

This packages aims to find, in a mixed exposure with many covariates and a future outcome, the exposure-covariate region combination that maximizes the differential impact of stochastic shift intervention. In a training fold we find the region in the covariate space that maximizes the average difference in stochastic shift interventions, compared between the regions. This outputs an exposure-covariate region pairing. Then in an estimation sample, we estimate a stochastic shift intervention using targeted learning in each level of the covariate, shifting the discovered exposure. This is done in a CV-TMLE procedure where each fold is used as validation and the complementary folds are used as training. This package outputs the targeted estimates for the discovered exposure in the levels of the discovered covariate both at the k-fold specific level and pooled across folds, which estimates the overall oracle parameter.

EffectXshift(
  w,
  a,
  y,
  deltas,
  estimator = c("tmle", "onestep"),
  fluctuation = c("standard", "weighted"),
  g_learner = NULL,
  mu_learner = NULL,
  n_folds = 2,
  outcome_type = "continuous",
  parallel = TRUE,
  parallel_type = "multi_session",
  num_cores = 2,
  seed = seed,
  hn_trunc_thresh = 10,
  adaptive_delta = FALSE,
  top_n = 1,
  min_obs = 20,
  density_classification = FALSE,
  max_depth = 1,
  rct = FALSE,
  rct_type = c("ate", "incps"),
  pval_thresh = 0.1
)

Arguments

w

A matrix, data.frame, or similar containing a set of baseline covariates. These variables are measured before exposures.

a

matrix, data.frame, or similar containing individual or multiple exposures.

y

numeric vector of observed outcomes.

deltas

A numeric value indicating the shift in exposures to define the target parameter, with respect to the scale of the exposures (A). If adaptive_delta is true, these values will be reduced.

estimator

The type of estimator to fit: "tmle" (default) for targeted maximum likelihood estimation, or "onestep" for a one-step estimator.

fluctuation

Method used in the targeting step for TML estimation: "standard" or "weighted". This determines where to place the auxiliary covariate in the logistic tilting regression.

g_learner

Learners for fitting the exposure mechanism g(A|W) via sl3. Used to estimate the density ratio for continuous exposures when density_classification = FALSE.

mu_learner

Learners for fitting Super Learner ensembles to the outcome model via sl3.

n_folds

Number of folds to use in cross-validation, default is 2.

outcome_type

Data type of the outcome, default is "continuous".

parallel

Whether to parallelize across cores (default: TRUE).

parallel_type

Type of parallelization to use if parallel is TRUE: "multi_session" (default), "multicore", or "sequential".

num_cores

Number of CPU cores to use in parallelization (default: 2).

seed

numeric seed value to be passed to all functions.

hn_trunc_thresh

Truncation level for the clever covariate (default: 10).

adaptive_delta

If TRUE, reduces the user-specified delta until the Hn calculated for a shift does not have any observation greater than hn_trunc_thresh (default: FALSE).

top_n

top number of effect modifier-exposure pairs to estimate.

min_obs

Minimum number of observations allowed in a covariate region.

density_classification

If TRUE, estimate the exposure density ratio via a classification reparameterization rather than direct conditional density estimation (default: FALSE). Continuous-exposure path only.

max_depth

Maximum depth of the partitioning tree used to discover effect-modification regions (default: 1).

rct

If TRUE, run the randomized-trial workflow for a single binary exposure (estimate the subject-level treatment effect and find the oracle effect-modification region). If FALSE (default), run the mixed-exposure stochastic-shift workflow.

rct_type

Only used when rct = TRUE. Either "ate" (default), which targets the subject-level ATE \(Q(1, W) - Q(0, W)\), or "incps", an incremental propensity-score shift from \(\alpha\) to \(\alpha + \delta\).

pval_thresh

p-value threshold for accepting a partition split when discovering effect-modification regions in the RCT workflow (default: 0.1).

Value

An S3 object of class EffectXshift containing the results of the procedure to compute a TML or one-step estimate of the counterfactual mean under a modified treatment policy that shifts a continuous-valued exposure by a scalar amount delta in subregions of the exposure space. These exposures are data-adaptively identified using the CV-TMLE procedure.