GSDCovAdj

The goal of GSDCovAdj is to combine group sequential, information-adaptive designs with covariate adjustment

Kelly Van Lancker, Josh Betz and Michael Rosenblum

Installation

Before installation of the development version of GSDCovAdj, we recommend installing the package simul from GitHub with:

# install.packages("devtools")
devtools::install_github("nt-williams/simul")

You can install the development version of GSDCovAdj from GitHub with:

devtools::install_github("kelvlanc/GSDCovAdj")

Scope

GSDCovAdj implements the methods proposed by Van Lancker, Betz and Rosenblum (2022) to combine group sequential, information-adaptive designs with covariate adjustment. The approach is implemented for a G-computation and targeted maximum likelihood estimator for continuous and binary outcomes, and for the estimator proposed by Díaz et al. (2019) for time-to-event outcomes.

Example

Installing R Packages

We first install the required packages

required_packages <-
  c("tidyr", "dplyr", "parallel", "ldbounds")

install.packages(required_packages)

Once the required packages are installed, they can be loaded using library()

library(rpact)
#> Warning: package 'rpact' was built under R version 4.1.2
library(tidyr)
#> Warning: package 'tidyr' was built under R version 4.1.2
library(dplyr)
#> Warning: package 'dplyr' was built under R version 4.1.2
#> 
#> Attaching package: 'dplyr'
#> The following objects are masked from 'package:stats':
#> 
#>     filter, lag
#> The following objects are masked from 'package:base':
#> 
#>     intersect, setdiff, setequal, union
library(parallel)
library(ldbounds)
#> Warning: package 'ldbounds' was built under R version 4.1.2

library(simul)
library(GSDCovAdj)
#> Warning: replacing previous import 'betareg::predict' by 'stats::predict' when
#> loading 'GSDCovAdj'

Design Parameters

We first determine the maximum/total information and sample size based on a design with 2 analyses: 1 interim analyses when 50% of the information is available in addition to the final analysis. The two-sided significance level equals 5% and the power to detect a 5% difference in the success rate (55% in treatment arm versus 50% in control) equals 90%. An alpha-spending function that approximates the Pocock boundaries are used for efficacy stopping (futility stopping is not considered).

design_par = getDesignGroupSequential(sided = 2, alpha = 0.05, beta=0.1,
                                      informationRates = c(0.50, 1),
                                      typeOfDesign = "asP")
# Determine Inflation Factor
#getDesignCharacteristics(design_par) #1.1110  

inf_total = 1.1110*(qnorm(0.975)+qnorm(0.90))^2/0.05^2

n_total = round(1.1110*(power.prop.test(power=0.90,p1=0.55,p2=0.5, 
                                         alternative="two.sided", 
                                         sig.level=0.05)$n)*2
)

We load the code to calculate expit.

expit = function(x){
  exp(x)/(1+exp(x))
}

Create Simulated Dataset

Create data frame ‘study_data’ with baseline covariates x_1, x_2, x_3 and x_4, a patient identifier (id), a treatment indicator (treatment), the primary outcome of interest (y_1), the time of enrollment (enrollment_times), and the time a patient’s outcome is measured (outcome_times).

set.seed(12345)
# Generate baseline covariates
baseline_covariates_orig <-
  matrix(
    data =
      mvtnorm::rmvnorm(
        n = n_total,
        mean = matrix(data = 0, nrow = 4),
        sigma = diag(x = 1, nrow = 4)
      ),
    nrow = n_total,
    ncol = 4
  ) %>% 
  data.frame() %>% 
  setNames(
    object = .,
    nm = paste0("x_", 1:4)
  )

# Generate treatment indicator
treatment <-
      rbinom(
        n = n_total,
        size = 1,
        prob = 0.5
      )

# (Counterfactual) outcomes under new experimental treatment
n_outcomes = 1
outcomes1 <-
      matrix(
        data =
          rbinom(
            n = n_total*n_outcomes,
            size = 1,
            prob = expit(-2.5*baseline_covariates_orig$x_1+
                           2*baseline_covariates_orig$x_2-
                           2.5*baseline_covariates_orig$x_3+
                           2.1*baseline_covariates_orig$x_4)
          ),
        nrow = n_total,
        ncol = n_outcomes
      ) %>% 
      data.frame() %>% 
      setNames(
        object = .,
        nm = paste0("y1_", 1:n_outcomes)
      )

# (Counterfactual) outcomes under control
outcomes0 <-
      matrix(
        data =
          rbinom(
            n = n_total*n_outcomes,
            size = 1,
            prob = expit(-2*baseline_covariates_orig$x_1+
                           2.5*baseline_covariates_orig$x_2-
                           2.25*baseline_covariates_orig$x_3-
                           2.1*baseline_covariates_orig$x_4)
          ),
        nrow = n_total,
        ncol = n_outcomes
      ) %>% 
      data.frame() %>% 
      setNames(
        object = .,
        nm = paste0("y0_", 1:n_outcomes)
      )

# Enrollment times
daily_enrollment = 5
enrollment_times <-
      round(
        runif(
          n = n_total, min = 0, max=n_total/daily_enrollment
        )
      )

# Outcome times
outcome_times <-
      setNames(
        object = 365 + enrollment_times,
        nm = paste0("outcome_time_", 1)
      )

# Measured baseline covariates   
baseline_covariates_meas = as.data.frame(cbind(
  x_1 = exp(baseline_covariates_orig$x_1/2),
  x_2 = baseline_covariates_orig$x_2/
  (1+exp(baseline_covariates_orig$x_1))+10,
  x_3 = (baseline_covariates_orig$x_1*
                                  baseline_covariates_orig$x_3/25+0.6)^3,
  x_4 = (baseline_covariates_orig$x_2
                                +baseline_covariates_orig$x_4+20)^2
  ))

# Make dataset    
study.data <-
      tibble(
        id = 1:n_total,
        baseline_covariates_meas,
        enrollment_times,
        treatment,
        outcome_times,
        y_1 = outcomes1$y1_1*treatment+outcomes0$y0_1*(1-treatment)
      )

The data can also be directly loaded from Github

data_url_gsd_data <-
  "https://github.com/jbetz-jhu/CovariateAdjustmentTutorial/raw/main/simData.RData"

load(file = url(data_url_gsd_data))

The complete simulated trial data without any missing values are in a data.frame named study.data.

Randomization Information
- treatment: Treatment Arm
Baseline Information
- x_1, x_2, x_3 and x_4: 4 continuous baseline covariates
- id: patient identifier
- enrollment_times: time of enrollment
Outcome information:
- y_1: Binary outcome (1: success; 0: otherwise)
- outcome_times: time someone’s outcome is measured (365 days after enrollment)

Combining Group Sequential Designs and Covariate Adjustment

Interim analysis

Via the function data_at_time_t() we construct a dataframe of the data available at the interim analysis, which is conducted when 50% of the patients have their primary endpoint available.

n_total_ia = round(0.5*n_total)
time_ia = sort(study.data$outcome_times)[n_total_ia]
n_recr_ia = length(which(study.data$enrollment_times<=time_ia))

data_ia = data_at_time_t(
  data = study.data,
  id_column = "id",
  enrollment_time = "enrollment_times",
  treatment_column = "treatment",
  covariate_columns = c("x_1", "x_2", "x_3", "x_4"),
  outcome_columns = c("y_1"),
  outcome_times = c("outcome_times"),
  analysis_time = time_ia
      )

We can then call the function interimAnalysis() to conduct the interim analysis. This gives us a decision based on the original estimator and updated estimator (which is the same as the original one at the first analysis). In addition, we get the estimates(both the original and updated estimate), the corresponding standard errors, test statistics, and information fractions. Finally, we also get the current covariance matrix based on the original estimates, which we will need at the final analysis to do the orthogonalization (in order to obtain the updated estimate).

results_ia = interimAnalysis(data = data_ia,
                             totalInformation = inf_total, 
                             estimationMethod= standardization,
                             estimand = "difference",
                             null.value = 0,
                             alpha = 0.05,
                             beta = 0.2,
                             alternative = "two.sided", 
                             typeOfDesign = "asP",
                             typeBetaSpending = "none",
                             futilityStopping = FALSE,
                             plannedAnalyses=2,
                             y0_formula=y_1 ~ x_1, 
                             y1_formula=y_1 ~ x_1,
                             family="binomial",
                             treatment_column = "treatment"
                             )

Results

Decision based on original interim estimator

results_ia$decisionOriginal
#> [1] "No"

So we do not stop the trial early for efficacy based on the original interim estimator.

Decision based on updated interim estimator

results_ia$decisionUpdated
#> [1] "No"

So we do not stop the trial early for efficacy based on the updated interim estimator.

Original interim estimate

results_ia$estimateOriginal
#> [1] 0.005660582
previousEstimates = results_ia$estimateOriginal

Updated interim estimate

results_ia$estimateUpdated
#> [1] 0.005660582

Covariance matrix (i.e., variance) of the original interim estimate

results_ia$covMatrixOriginal
#> [1] 0.0003517528
covMatrix = results_ia$covMatrixOriginal

Information time (current information divided by expected information at end of trial) corresonding with original interim estimate

results_ia$informationTimeOriginal
#> [1] 0.6088245
previousTimes = results_ia$informationTimeOriginal

Information time (current information divided by expected information at end of trial) corresonding with updated interim estimate

results_ia$informationTimeUpdated
#> [1] 0.6088245
previousTimesUpd = results_ia$informationTimeUpdated

Final analysis

Via the function data_at_time_t() we construct a dataframe of the data available at the final analysis (which equals the full dataset!).

time_fin = sort(study.data$outcome_times)[n_total]
n_recr_fin = length(which(study.data$enrollment_times<=time_fin))

data_fin = data_at_time_t(
  data = study.data,
  id_column = "id",
  enrollment_time = "enrollment_times",
  treatment_column = "treatment",
  covariate_columns = c("x_1", "x_2", "x_3", "x_4"),
  outcome_columns = c("y_1"),
  outcome_times = c("outcome_times"),
  analysis_time = time_fin
      )

We can then call the function interimAnalysis() to conduct the final analysis. This gives us a decision based on the original estimator and updated estimator (which is the same as the original one at the first analysis). In addition, we get the estimates(both the original and updated estimate), the corresponding standard errors, test statistics, and information fractions. Finally, we also get the current covariance matrix based on the original estimates, which we will need at the final analysis to do the orthogonalization (in order to obtain the updated estimate).

results_fin = interimAnalysis(data = data_fin, 
                            totalInformation = inf_total, 
                            estimationMethod= standardization,
                            estimand = "difference",
                            previousEstimatesOriginal=previousEstimates,
                            previousCovMatrixOriginal=covMatrix,
                            previousInformationTimesOriginal=previousTimes,
                            previousInformationTimesUpdated=previousTimesUpd,
                            previousDatasets = list(data_ia),
                            null.value = 0,
                            alpha = 0.05, 
                            beta = 0.2,
                            alternative = "two.sided", 
                            typeOfDesign = "asP",
                            typeBetaSpending = "none",
                            futilityStopping = FALSE,
                            plannedAnalyses=2,
                            y0_formula=y_1 ~ x_2, 
                            y1_formula=y_1 ~ x_2,
                            family="binomial",
                            parametersPreviousEstimators = list(list(
                              y0_formula=y_1 ~ x_1,
                              y1_formula=y_1 ~ x_1,
                              family="binomial"
                              )),
                            treatment_column = "treatment"
)
#> Warning: Argument unknown in getDesignGroupSequential(...): 'y0_formula' =
#> formula will be ignored
#> Warning: Argument unknown in getDesignGroupSequential(...): 'y1_formula' =
#> formula will be ignored
#> Warning: Argument unknown in getDesignGroupSequential(...): 'family' =
#> 'binomial' will be ignored
#> Warning: Argument unknown in getDesignGroupSequential(...): 'treatment_column' =
#> 'treatment' will be ignored
#> Warning: Argument unknown in getDesignGroupSequential(...): 'y0_formula' =
#> formula will be ignored
#> Warning: Argument unknown in getDesignGroupSequential(...): 'y1_formula' =
#> formula will be ignored
#> Warning: Argument unknown in getDesignGroupSequential(...): 'family' =
#> 'binomial' will be ignored
#> Warning: Argument unknown in getDesignGroupSequential(...): 'treatment_column' =
#> 'treatment' will be ignored

Results

Decision based on original final estimator

results_fin$decisionOriginal
#>   X2 
#> "No"

So we do not reject the null hypothesis based on the original final estimator.

Decision based on updated final estimator

results_fin$decisionUpdated
#>      [,1]
#> [1,] "No"

So we do not reject the null hypothesis based on the updated final estimator.

Original final estimate

results_fin$estimateOriginal
#> [1] 0.01689733

Updated final estimate

results_fin$estimateUpdated
#>           [,1]
#> [1,] 0.0152497

statneutrino / gsdcovadj Goto Github PK

gsdcovadj's Introduction

GSDCovAdj

Installation

Scope

Example

Installing R Packages

Design Parameters

Create Simulated Dataset

Combining Group Sequential Designs and Covariate Adjustment

Interim analysis

Results

Final analysis

Results

gsdcovadj's People

Contributors

Recommend Projects

React

Vue.js

Typescript

TensorFlow

Django

Laravel

D3

Recommend Topics

javascript

web

server

Machine learning

Visualization

Game

Recommend Org

Facebook

Microsoft

Google

Alibaba

D3

Tencent