TargetedLearning.CTMLEs

Exported

Performs colaborative targeted minimum loss-based estimation

Arguments

logitQnA1 - Vector of length n of initial estimates logit(ar{Q}_n(1, W_i))
logitQnA0 - Vector of length n of initial estimates logit(ar{Q}_n(0, W_i))
W - Matrix of covariates to be potentially used to estimate g. n rows. The first column should be all ones for an intercept.
A - Vector of length n of treatments, 0.0 or 1.0
Y - Vector of length n of outcomes, bounded between 0 and 1.

Keyword Arguments

param - Target parameter. See ?ATE or ?Mean. Default is ATE()
gbounds - values for bounding g. Estimates of g will be bounded by extrema(gbounds). To only bound away from 0 (or 1), set to e.g. [0.01, 1.0] (or [0.0, 0.99]). To turn off bounding all together, set to [0.0, 1.0]. Defaults to [0.01, 0.99].
searchstrategy - Strategy for adding covariates to estimates of g. See ?SearchStrategy. Default is ForwardStepwise()
cvplan - An iterator of vectors of Int indexes of observations in each training set for cross validaiton. StratifiedKfold and StratifiedRandomSub from the MLBase package are useful here. Defaults to 10 fold CV stratifying by treatment for outcomes with more than 2 levels, or by treatment and outcome for binary outcomes.
patience - For how many steps should CV continue after finding a local optimum? Defaults to typemax(Int)

Defines how the next covariate added to g is chosen

ForwardStepwise - default, tries each unused covariate to find the one that helps the most. Slow, O(p^2) time for p covariates.
PreOrdered - Orders covariates once then adds them in that order. Takes O(p) time for p covariates. Specify with PreOrdered(ordering) where ordering is
- LogisticOrdering() - ranks covariates based on reduction in log likelihood with intital Q as an offset if added one at a time. Basically the first step of ForwardStepwise
- PartialCorrOrdering() - ranks covariates based on partial correlation of each covariate and the residual (y - predicted y from initial Qbar) given a
- HDPSOrdering() - ranks covariates based on bias reduction potential as in the hdps procedure. This ignores intial Qbar.
- ManualOrdering(ord) - orders covars based on specified indexes, a Vector{Int}, of column indexes of W, not including 1, the intercept.