Help for package crossurr

Type:

Package

Title:

Cross-Fitting for Doubly Robust Evaluation of High-Dimensional Surrogate Markers

Version:

1.1.2

Description:

Doubly robust methods for evaluating surrogate markers as outlined in: Agniel D, Hejblum BP, Thiebaut R & Parast L (2022). "Doubly robust evaluation of high-dimensional surrogate markers", Biostatistics <doi:10.1093/biostatistics/kxac020>. You can use these methods to determine how much of the overall treatment effect is explained by a (possibly high-dimensional) set of surrogate markers.

License:

MIT + file LICENSE

Depends:

R (≥ 3.6.0)

Imports:

dplyr, gbm, glmnet, glue, parallel, pbapply, purrr, ranger, RCAL, rlang, SIS, stats, SuperLearner, tibble, tidyr

Encoding:

UTF-8

RoxygenNote:

7.3.2

NeedsCompilation:

Packaged:

2025-04-04 12:54:18 UTC; boris

Author:

Denis Agniel [aut, cre], Boris P. Hejblum [aut], Layla Parast [aut]

Maintainer:

Denis Agniel <dagniel@rand.org>

Repository:

CRAN

Date/Publication:

2025-04-08 13:50:02 UTC

crossurr

Description

The main functions of this package are xf_surrogate and xfr_surrogate

Author(s)

Maintainer: Denis Agniel dagniel@rand.org

Authors:

Boris P. Hejblum boris.hejblum@u-bordeaux.fr
Layla Parast parast@austin.utexas.edu

lasso

Description

lasso

Usage

lasso(
  x = NULL,
  y = NULL,
  data = NULL,
  newX = NULL,
  newX0 = NULL,
  newX1 = NULL,
  relax = TRUE,
  ps_fit = FALSE,
  ...
)

Ordinary Least Squares

Description

Ordinary Least Squares

Usage

ols(
  x = NULL,
  y = NULL,
  data = NULL,
  test_data = NULL,
  test_data0 = NULL,
  test_data1 = NULL,
  ...
)

A simple function to simulate example data.

Description

A simple function to simulate example data.

Usage

sim_data(n, p)

Arguments

n

number of simulated observations

p

number of simulated variables

Value

toy dataset used for demonstrating the methods with outcome y, treatment a, covariates x.1, x.2, and surrogates s.1, s.2, ...

A function for estimating the proportion of treatment effect explained using cross-fitting.

Description

A function for estimating the proportion of treatment effect explained using cross-fitting.

Usage

xf_surrogate(
  ds,
  x = NULL,
  s,
  y,
  a,
  K = 5,
  outcome_learners = NULL,
  ps_learners = outcome_learners,
  interaction_model = TRUE,
  trim_at = 0.05,
  outcome_family = gaussian(),
  mthd = "superlearner",
  n_ptb = 0,
  ncores = parallel::detectCores() - 1,
  ...
)

Arguments

ds

a data.frame.

x

names of all covariates in ds that should be included to control for confounding (eg. age, sex, etc). Default is NULL.

s

names of surrogates in ds.

y

name of the outcome in ds.

a

treatment variable name (eg. groups). Expect a binary variable made of 1s and 0s.

K

number of folds for cross-fitting. Default is 5.

outcome_learners

string vector indicating learners to be used for estimation of the outcome function (e.g., "SL.ridge"). See the SuperLearner package for details.

ps_learners

string vector indicating learners to be used for estimation of the propensity score function (e.g., "SL.ridge"). See the SuperLearner package for details.

interaction_model

logical indicating whether outcome functions for treated and control should be estimated separately. Default is TRUE.

trim_at

threshold at which to trim propensity scores. Default is 0.05.

outcome_family

default is 'gaussian' for continuous outcomes. Other choice is 'binomial' for binary outcomes.

mthd

selected regression method. Default is 'superlearner', which uses the SuperLearner package for estimation. Other choices include 'lasso' (which uses glmnet), 'sis' (which uses SIS), 'cal' (which uses RCAL).

n_ptb

Number of perturbations. Default is 0 which means asymptotic standard errors are used.

ncores

number of CPUs used for parallel computations. Default is parallel::detectCores()-1

...

additional parameters (in particular for super_learner)

Value

a tibble with columns:

R: estimate of the proportion of treatment effect explained, equal to 1 - deltahat_s/deltahat.
R_se standard error for the PTE.
deltahat_s: residual treatment effect estimate.
deltahat_s_se: standard error for the residual treatment effect.
pi_o: estimate of the proportion of overlap.
R_o: PTE only in the overlap region.
R_o_se: the standard error for R_o.
deltahat_s_o: residual treatment effect in overlap region,
deltahat_s_se_o: standard error for deltahat_s_o.
deltahat: overall treatment effect estimate.
deltahat_se: standard error for overall treatment effect estimate.
delta_diff: difference between the treatment effects, equal to the numerator of PTE.
dd_se: standard error for delta_diff

Examples


n <- 300
p <- 50
q <- 2
wds <- sim_data(n = n, p = p)

if(interactive()){
 sl_est <- xf_surrogate(ds = wds,
   x = paste('x.', 1:q, sep =''),
   s = paste('s.', 1:p, sep =''),
   a = 'a',
   y = 'y',
   K = 4,
   trim_at = 0.01,
   mthd = 'superlearner',
   outcome_learners = c("SL.mean","SL.lm", "SL.svm", "SL.ridge"),
   ps_learners = c("SL.mean", "SL.glm", "SL.svm", "SL.lda"),
   ncores = 1)

 lasso_est <- xf_surrogate(ds = wds,
   x = paste('x.', 1:q, sep =''),
   s = paste('s.', 1:p, sep =''),
   a = 'a',
   y = 'y',
   K = 4,
   trim_at = 0.01,
   mthd = 'lasso',
   ncores = 1)
}

Title

Description

Title

Usage

xfit_dr(
  ds,
  x,
  y,
  a,
  K = 5,
  outcome_learners = NULL,
  ps_learners = outcome_learners,
  interaction_model = TRUE,
  trim_at = 0.05,
  outcome_family = gaussian(),
  mthd = "superlearner",
  ncores = parallel::detectCores() - 1,
  ...
)

A function for estimating the proportion of treatment effect explained using repeated cross-fitting.

Description

A function for estimating the proportion of treatment effect explained using repeated cross-fitting.

Usage

xfr_surrogate(
  ds,
  x = NULL,
  s,
  y,
  a,
  splits = 50,
  K = 5,
  outcome_learners = NULL,
  ps_learners = NULL,
  interaction_model = TRUE,
  trim_at = 0.05,
  outcome_family = gaussian(),
  mthd = "superlearner",
  n_ptb = 0,
  ...
)

Arguments

ds

a data.frame.

x

names of all covariates in ds that should be included to control for confounding (eg. age, sex, etc). Default is NULL.

s

names of surrogates in ds.

y

name of the outcome in ds.

a

treatment variable name (eg. groups). Expect a binary variable made of 1s and 0s.

splits

number of data splits to perform.

K

number of folds for cross-fitting. Default is 5.

outcome_learners

string vector indicating learners to be used for estimation of the outcome function (e.g., "SL.ridge"). See the SuperLearner package for details.

ps_learners

string vector indicating learners to be used for estimation of the propensity score function (e.g., "SL.ridge"). See the SuperLearner package for details.

interaction_model

logical indicating whether outcome functions for treated and control should be estimated separately. Default is TRUE.

trim_at

threshold at which to trim propensity scores. Default is 0.05.

outcome_family

default is 'gaussian' for continuous outcomes. Other choice is 'binomial' for binary outcomes.

mthd

n_ptb

Number of perturbations. Default is 0 which means asymptotic standard errors are used.

...

additional parameters (in particular for super_learner)

Value

a tibble with columns:

Rm: estimate of the proportion of treatment effect explained, computed as the median over the repeated splits.
R_se0 standard error for the PTE, accounting for the variability due to splitting.
R_cil0 lower confidence interval value for the PTE.
R_cih0 upper confidence interval value for the PTE.
Dm: estimate of the overall treatment effect, computed as the median over the repeated splits.
D_se0 standard error for the overall treatment effect, accounting for the variability due to splitting.
D_cil0 lower confidence interval value for the overall treatment effect.
D_cih0 upper confidence interval value for the overall treatment effect.
Dsm: estimate of the residual treatment effect, computed as the median over the repeated splits.
Ds_se0 standard error for the residual treatment effect, accounting for the variability due to splitting.
Ds_cil0 lower confidence interval value for the residual treatment effect.
Ds_cih0 upper confidence interval value for the residual treatment effect.

Examples


n <- 100
p <- 20
q <- 2
wds <- sim_data(n = n, p = p)

if(interactive()){
 lasso_est <- xfr_surrogate(ds = wds,
   x = paste('x.', 1:q, sep =''),
   s = paste('s.', 1:p, sep =''),
   a = 'a',
   y = 'y',
   splits = 2,
   K = 2,
   trim_at = 0.01,
   mthd = 'lasso',
   ncores = 1)
}