R/pc_power_fast.r

In CrumpLab/conflictPower: Simulation based power analysis for adaptive control designs

#' Estimate power for proportion conflict effect using monte-carlo simulation
#'
#' @param subjects A number for the number of subjects in simulated experiment
#' @param mc_c_nmst A vector for mostly conflict (mc), conflict trials (c) containing the parameters for an ex-gaussian distribution, c(n, mu, sigma, tau), where n is the number of trials.
#' @param mc_nc_nmst A vector for mostly conflict (mc), conflict trials (c) containing the parameters for an ex-gaussian distribution, c(n, mu, sigma, tau), where n is the number of trials.
#' @param mnc_c_nmst A vector for mostly no-conflict (mnc), conflict trials (c) containing the parameters for an ex-gaussian distribution, c(n, mu, sigma, tau), where n is the number of trials.
#' @param mnc_nc_nmst A vector for mostly no-conflict (mnc), conflict trials (c) containing the parameters for an ex-gaussian distribution, c(n, mu, sigma, tau), where n is the number of trials.
#' @param num_sims A number, simulations to run
#' @param alpha A number, alpha criterion
#' @return A number, power: the proportion of simulated experiments that returned a p-value less than the alpha criterion for the conflict effect
#' @details  This function uses monte-carlo simulation to determine statistical power associated for detecting a proportion conflict effect, specifically a difference between two conflict effect, typically based on a manipulation of the proportion of conflict and no-conflict trials.
#'
#' For every simulated experiment, a one sample t-test  (two-tailed) is computed, and the p-value is saved. Power is the proportion of simulated experiments that return p-values less than the defined alpha criterion.
#'
#' @examples
#' pc_power_fast(subjects=10,
#'   mc_c_nmst = c(80,550,100,100),
#'   mc_nc_nmst = c(20,500,100,100),
#'   mnc_c_nmst = c(20,570,100,100),
#'   mnc_nc_nmst = c(80,500,100,100),
#'   num_sims = 1000,
#'   alpha = .05)
#' @export

pc_power_fast <- function(subjects,
                            mc_c_nmst,
                            mc_nc_nmst,
                            mnc_c_nmst,
                            mnc_nc_nmst,
                            num_sims=1000,
                            alpha=.05) {

  p_values <- c()
  for(i in 1: num_sims){
    mc_c_mat <- matrix(rexgauss(subjects*mc_c_nmst[1],
                                      mc_c_nmst[2],mc_c_nmst[3],mc_c_nmst[4],
                                      positive=TRUE),
                    nrow=subjects,ncol=mc_c_nmst[1])
    mc_nc_mat <- matrix(rexgauss(subjects*mc_nc_nmst[1],
                                       mc_nc_nmst[2],mc_nc_nmst[3],mc_nc_nmst[4],
                                       positive=TRUE),
                     nrow=subjects,ncol=mc_nc_nmst[1])
    mnc_c_mat <- matrix(rexgauss(subjects*mnc_c_nmst[1],
                                         mnc_c_nmst[2],mnc_c_nmst[3],mnc_c_nmst[4],
                                         positive=TRUE),
                       nrow=subjects,ncol=mnc_c_nmst[1])
    mnc_nc_mat <- matrix(rexgauss(subjects*mnc_nc_nmst[1],
                                          mnc_nc_nmst[2],mnc_nc_nmst[3],mnc_nc_nmst[4],
                                          positive=TRUE),
                        nrow=subjects,ncol=mnc_nc_nmst[1])
    diff_mat <- (rowMeans(mnc_c_mat)-rowMeans(mnc_nc_mat))-(rowMeans(mc_c_mat)-rowMeans(mc_nc_mat))
    p_values[i] <- t.test(diff_mat)$p.value
  }

  power <- length(p_values[p_values < alpha])/num_sims
  return(power)

}