BETA: Covariate adjustment using working models from the super learner libraries through the AIPW package with cross-fitting.

Estimate treatment-group-specific response means and (optionally) treatment group contrasts using a generalized linear working model.

robincar_SL(
  df,
  treat_col,
  response_col,
  car_strata_cols = NULL,
  covariate_cols = NULL,
  car_scheme = "simple",
  covariate_to_include_strata = NULL,
  SL_libraries = c(),
  SL_learners = c(),
  k_split = 2,
  g_accuracy = 7,
  contrast_h = NULL,
  contrast_dh = NULL,
  variance_type = 1
)

Arguments

df

A data.frame with the required columns

treat_col

Name of column in df with treatment variable

response_col

Name of the column in df with response variable

car_strata_cols

Names of columns in df with car_strata variables

covariate_cols

Names of columns in df with covariate variables

car_scheme

Name of the type of covariate-adaptive randomization scheme. One of: "simple", "pocock-simon", "biased-coin", "permuted-block".

covariate_to_include_strata

Whether to include car_strata variables in covariate adjustment. Defaults to F for ANOVA and ANCOVA; defaults to T for ANHECOVA. User may override by passing in this argument.

SL_libraries

Vector of super-learner libraries to use for the covariate adjustment (see SuperLearner::listWrappers)

SL_learners

Optional list of super-learner "learners" to use for the covariate adjustment (see SuperLearner::create.Learner())

k_split

Number of splits to use in cross-fitting

g_accuracy

Level of accuracy to check prediction un-biasedness (in digits).

contrast_h

An optional function to specify a desired contrast

contrast_dh

An optional jacobian function for the contrast (otherwise use numerical derivative)

variance_type

The type of variance estimator to use, type 1, 2, or 3. All three are asymptotically equivalent. See details of RobinCar::robincar_glm for more.

Value

See value of RobinCar::robincar_glm, but the working model for \(\hat{\mu}(X_i)\) is based on the AIPW::AIPW package that uses specified SuperLearner libraries and cross-fitting. Also, `mod` attribute is an object of class AIPW::AIPW.

Details

*WARNING: This function is still under development and has not been extensively tested.* This function currently only works for two treatment groups. Before using this function, you must load the SuperLearner library with `library(SuperLearner)`, otherwise the function call will fail.

Examples

library(SuperLearner)
#> Loading required package: nnls
#> Loading required package: gam
#> Loading required package: splines
#> Loading required package: foreach
#> Loaded gam 1.22-3
#> Super Learner
#> Version: 2.0-29
#> Package created on 2024-02-06
library(ranger)
n <- 1000
set.seed(10)
DATA2 <- data.frame(A=rbinom(n, size=1, prob=0.5),
                    y=rbinom(n, size=1, prob=0.2),
                    x1=rnorm(n),
                    x2=rnorm(n),
                    x3=as.factor(rbinom(n, size=1, prob=0.5)),
                    z1=rbinom(n, size=1, prob=0.5),
                    z2=rbinom(n, size=1, prob=0.5))
DATA2[, "y"] <- NA
As <- DATA2$A == 1
DATA2[DATA2$A == 1, "y"] <- rbinom(
  sum(As),
  size=1,
  prob=exp(DATA2[As,]$x1)/(1+exp(DATA2[As,]$x1)))
DATA2[DATA2$A == 0, "y"] <- rbinom(
  n-sum(As),
  size=1,
  prob=exp(1 +
    5*DATA2[!As,]$x1 + DATA2[!As,]$x2)/
    (1+exp(1 + 5*DATA2[!As,]$x1 + DATA2[!As,]$x2)))
DATA2$A <- as.factor(DATA2$A)

sl.mod <- robincar_SL(
  df=DATA2,
  response_col="y",
  treat_col="A",
  car_strata_cols=c("z1"),
  covariate_cols=c("x1"),
  SL_libraries=c("SL.ranger"),
  car_scheme="permuted-block",
  covariate_to_include_strata=TRUE
)
#> Done!
#> Warning: Prediction unbiasedness does not hold.

sl.mod$result
#> # A tibble: 2 × 4
#>   treat estimate     se `pval (2-sided)`
#>   <chr>    <dbl>  <dbl>            <dbl>
#> 1 0        0.582 0.0194        2.15e-198
#> 2 1        0.518 0.0217        2.85e-126