Estimate the Outcome Mechanism

Estimate the Outcome Mechanism

Usage

est_Q(
  Y,
  C_cens = rep(1, length(Y)),
  A,
  W,
  delta = 0,
  samp_weights = rep(1, length(Y)),
  fit_type = c("sl", "glm"),
  glm_formula = "Y ~ .",
  sl_learners = NULL,
  scale_outcome = TRUE
)

Arguments

Y

A numeric vector of observed outcomes.

C_cens

A numeric vector of loss to follow-up indicators.

A

A numeric vector of observed exposure values.

W

A numeric matrix of observed baseline covariate values.

delta

A numeric indicating the magnitude of the shift to be computed for the exposure A. This is passed to the internal shift_additive and is currently limited to additive shifts.

samp_weights

A numeric vector of observation-level sampling weights, as produced by the internal procedure to estimate the two-phase sampling mechanism est_samp.

fit_type

A character indicating whether to use GLMs or Super Learner to fit the outcome regression. If the option "glm" is selected, the argument glm_formula must NOT be NULL, instead containing a model formula (as per glm) as a character. If the option "sl" is selected, the argument sl_learners must NOT be NULL; instead, an instantiated sl3 Lrnr_sl object, specifying learners and a metalearner for the Super Learner fit, must be provided. Consult the documentation of sl3 for details.

glm_formula

A character giving a formula for fitting a (generalized) linear model via glm.

sl_learners

Object containing a set of instantiated learners from the sl3, to be used in fitting an ensemble model.

scale_outcome

A logical indicating whether the outcome should be scaled to the unit interval before estimating the outcome regression (the default). If FALSE, the outcome regression is estimated on the original scale of the outcome (e.g., for count outcomes, with learners that require non-negative integers), and its predictions are then scaled to the unit interval, as required by the targeting step. When FALSE, GLM-based fits use a Gaussian family; other working models (e.g., Poisson) may be specified via sl_learners.

Value

A data.table with two columns, containing estimates of the outcome mechanism at the natural value of the exposure Q(A, W) and an upshift of the exposure Q(A + delta, W).

Details

Compute the outcome regression for the observed data, including with the shift imposed by the intervention. This returns the outcome regression for the observed data (at A) and under the counterfactual shift shift (at A + delta).