EYdopt: Estimation of E[Ydopt]

In lmmontoya/SL.ODTR: SuperLearner Estimation and Evaluation of ODTR

Estimation of E[Ydopt]

Description

Given a W, A, Y dataset, this function will compute the estimated ODTR using SuperLearner. If a Qbar function is provided that computes the true E[Y|A,W] (e.g., if simulating), the function will also return the true treatment under the optimal rule and other metrics of evaluating the estimated optimal rule's performance. Then, it will estimate E[Ydopt] using g-computation, IPTW, IPTW-DR, TMLE, and CV-TMLE. Follows the framework of Luedtke and van der laan, 2015 and 2016.

Usage

EYdopt(
  W,
  V,
  A,
  Y,
  g.SL.library = "SL.mean",
  QAW.SL.library,
  blip.SL.library,
  dopt.SL.library = NULL,
  metalearner = "blip",
  risk.type = "CV TMLE",
  grid.size = 100,
  VFolds = 10,
  QAW.fun = NULL,
  family = NULL,
  contrast = NULL
)

Arguments

W	Data frame of observed baseline covariates
V	Data frame of observed baseline covariates (subset of W) used to design the ODTR
A	Vector of treatment
Y	Vector of outcome (continuous or binary)
g.SL.library	SuperLearner library for estimating txt mechanism
QAW.SL.library	SuperLearner library for estimating outcome regression
blip.SL.library	SuperLearner library for estimating the blip
dopt.SL.library	SuperLearner library for estimating dopt directly. Default is NULL. Could be "DonV", "Qlearn", "OWL", "EARL", "optclass", "RWL", "treatall", "treatnone". Could also be "all" for all algorithms.
metalearner	Discrete ("discrete"), blip-based ("blip"), vote-based SuperLearner ("vote"). Note that if metalearner is "vote" then cannot put in kappa.
risk.type	Risk type in order to pick optimal combination of coefficients to combine the candidate algorithms. For (1) MSE risk use "CV MSE"; for (2) -E[Ydopt] risk use "CV IPCWDR" (for -E[Ydopt] estimated using double-robust IPTW) or "CV TMLE" (for -E[Ydopt] estimates using TMLE); (3) For the upper bound of the CI of -E[Ydopt] use "CV TMLE CI"
grid.size	Grid size for simplex.sample() function to create possible combinations of coefficients
VFolds	Number of folds to use in cross-validation. Default is 10.
QAW.fun	True outcome regression E[Y|A,W]. Useful for simulations. Default is NULL.
family	either "gaussian" or "binomial". Default is null, if outcome is between 0 and 1 it will change to binomial, otherwise gaussian
contrast	A dim = (n, num contrasts) matrix or dataframe (with columns preferably named) to contrast Psi = E[Ydopt]-E[Ycontrast] for CV-TMLE. For example, contrast = data.frame("EY0" = rep(0,n)) will contrast Psi = E[Ydopt]-E[Y0]. Default is NULL.
odtr.obj	An object from the odtr function that estimates the odtr.

Value

If the true Qbar function is specified, the output will be a vector of point estimates of E[Ydopt] and their respective confidence intervals. This will be for both the estimated optimal rule and the true optimal rule. Performance results on the optimal rule will also be output: proportion of people treated under ODTR, proportion of times the estimated rule matches the optimal rule, the mean outcome under the estimated optimal rule under the true mean outcome function, and the mean outcome under the estimated optimal rule under the sample-specific true mean outcome.

If the true Qbar is not specified, return:

EYdopt_estimates

Point estimates and confidence intervals for E[Ydopt], using the unadjusted mean outcome for the people who received the optimal rule, g-computation, IPTW, IPTW-DR, TMLE

SL.odtr

SuperLearner list. See SL.blip or SL.vote documentation.

References

van der Laan, Mark J., and Alexander R. Luedtke. "Targeted learning of the mean outcome under an optimal dynamic treatment rule." Journal of causal inference 3.1 (2015): 61-95.

Luedtke, Alexander R., and Mark J. van der Laan. "Super-learning of an optimal dynamic treatment rule." The international journal of biostatistics 12.1 (2016): 305-332.

Luedtke, Alexander R., and Mark J. van der Laan. "Optimal individualized treatments in resource-limited settings." The international journal of biostatistics 12.1 (2016): 283-303.

Coyle, J.R. (2017). Jeremy Coyle, “Computational Considerations for Targeted Learning” PhD diss., University of California, Berkeley 2017 https://escholarship.org/uc/item/9kh0b9vm.

Examples

## Example
library(SuperLearner)
library(hitandrun)
ObsData = subset(DGP_bin_simple(1000), select = -c(A_star, Y_star))
W = subset(ObsData, select = -c(A,Y))
V = W
A = ObsData$A
Y = ObsData$Y

# E[Ydopt] using blip-based estimate of ODTR with risk function CV-TMLE
EYdopt(W = W, A = A, Y = Y, V = W, blip.SL.library = "SL.blip.HTEepi", g.SL.library = "SL.mean", QAW.SL.library = "SL.QAW.HTEepi", risk.type = "CV TMLE", metalearner = 'blip')

lmmontoya/SL.ODTR documentation built on Feb. 24, 2024, 10:46 a.m.