est_effect: Efficient Estimation of Causal Effects Under Dynamic...
In Netflix/sherlock: Causal Machine Learning for Segment Discovery and Analysis

Compute efficient estimates of the heterogeneous treatment effect (HTE) or the optimal treatment effect (OTE), as well as related parameters, including subgroup average treatment effects and the counterfactual mean of assigning the dynamic treatment rule.

1 2	est_effect(data_with_rule, param_type = c("hte", "ote"), est_type = c("onestep", "tmle"))

data_with_rule

A data.table containing the input data, augmented with cross-validated nuisance parameter estimates, an estimate of the CATE, and a treatment rule assigned based on the estimated CATE via assign_rule. This input object should be created by successive calls to set_est_data, est_cate, and assign_rule in a sequence, or through a wrapper function that composes these function calls automatically.

param_type

A character providing a specification for a family of target parameters to be estimated. The two choices provide estimates of a range of target parameters. Specifically, the choices correspond to

"hte", which computes the average treatment effect (ATE), which contrasts static interventions for treatment assignment and treatment being withhold, for (1) the full population, (2) within the subgroup of units dynamically assigned treatment, and (3) within the subgroup of units from whom treatment was withheld dynamically. Also included is the heterogeneous treatment effect (HTE), defined as a difference of the two subgroup ATEs, which captures the benefit attributable to treating those units who should receive treatment and withholding treatment from those that could be harmed by the treatment.
"ote", which computes several counterfactual means, for (1) the static intervention of assigning treatment to all units, (2) the static intervention of withholding treatment from all units, (3) the dynamic intervention of assigning treatment to those predicted to benefit from it while withholding treatment from those units that could be harmed. In addition, two average treatment effects are evaluated, each contrasting the dynamic treatment rule against the static interventions of assigning or withholding treatment. The latter three parameters capture contrasts based on the optimal treatment effect (OTE).

est_type

A character specifying the type of estimator to be computed. Both estimators are asymptotically linear when flexible modeling choices are used to estimate nuisance parameters, doubly robust (consistent when at least one nuisance parameter is correctly estimated), and achieve the best possible variance (i.e., asymptotically efficient) among the class of regular asymptotically linear estimators. The two options are

"onestep", corresponding to the one-step estimator, a first-order solution to the efficient influence function (EIF) estimating equation. This is not a substitution (direct) estimator and may be unstable in the sense of yielding estimates outside the bounds of the target parameter.
"tmle", corresponding to the targeted minimum loss estimator, an approach that updates initial estimates of the outcome model by way of a one-dimensional fluctuation model that aims to approximately solve the EIF estimating equation.

A data.table with a handful of rows, one for each target parameter estimated, and columns giving the parameter name the point estimate of the target parameter, and the standard error of the estimate. Also included are the lower and upper confidence limits for the point estimated based on Wald-style confidence intervals.

Netflix/sherlock documentation built on Dec. 17, 2021, 5:22 a.m.