keilsim: Function for running simulations comparing bayesian and...

In stablemarkets/KeilSim: Function for performing simulations behind Statistical Computing Final Report

Description

Simulates causal effects under setting detailed in section 4.2 of Keil et al. 2017 paper "A Bayesian approach to the g-formula". Simulation can be conducted under varying sample sizes, n, varying true risk difference values, RD, and either under a correct model, misspecified=FALSE, or misspecified model, misspecified=TRUE.

Usage

keilsim(n = 20, RD = 0, N_sims = 1000, mcmc_iter = 10000,
  warmup_iter = 9900, N_gcomp = 1000, boot_iter = 100,
  output_all = FALSE, misspecified = FALSE, parallel = FALSE,
  ncores = NULL)

Arguments

n	scalar, positive interger greater than 1. This is the sample size in each simulate dataset.
RD	scalar, numeric values. This represents the true risk difference under which data are simulated.
N_sims	scalar, positive interger for number of simulated data sets to use.
mcmc_iter	scalar, positive interger for total number of MCMC draws to take from the posterior when performing Bayesian g-compuation. Should have at least 1000 posterior draws after a sufficient warm-up period.
warmup_iter	scalar, positive interger for number of warm-up (aka burn-in) draws when performing Bayesian g-computation. Must be < mcmc_iter. The total number of draws used is mcmc_iter - warmup_iter. Recommended to use at least 1000 warm-up.
N_gcomp	scalar, positive interger for number of MCMC iterations to use when performing the integral involved in g-computation. See Details.
boot_iter	number of nonparametric bootstrap resamples to use when performing frequentist g-computation. See Details.
output_all	logical (TRUE/FALSE). If TRUE, outputs estimate for each simulated dataset. If FALSE, just outputs summary statistics across all simulated datasets.
misspecified	logical (TRUE/FALSE). If TRUE, both frequentist and Bayesian g-computation is performing without adjusting for confounding. If FALSE, both models correctly adjust for confounding.
parallel	logical (TRUE/FALSE). If TRUE, parallel processing is used to perform N_sims simulates in parallel. If FALSE, only single core is used. See Details.
ncores	scalar, positive interger for number of cores to use if parallel==TRUE. Must be between 1 and max cores. Use snow::detectCores() to find the maximum available cores on your machine.

Details

keilsim() performs the simulation described in Keil et al 2017 for desired simulation settings. In each iteration, we perform a Bayesian g-computation and a frequentist g-computation. Bayesian models are estimated using STAN (which in turn back-ends to C++) in the back end. R's glm() is used to estimate frequentist models. Both g-computations use MCMC to evaluate the integral involved in g-computation. The number of MCMC iterations to use in this integration is gcomp_iter. Since the data generation process involves only two time periods and binary treatments and confounders, we only need about 100 g-computation interations to accurately estimate the integral. For the frequentist method, an interval estimate for the causal effect is calculated using nonparametric bootstrap. boot_iter resamples with replacement are used. Percentiles of the sampling distribution are used to form intervals. The function includes an option to run the N_sims simulations in parallel. This is enabled for both Macs and PCs by implementing doParallel, however **reproducibility is NOT gauranteed when running in parallel** since set.seed() is not respected. set.seed() is respected only when not running in parallel.

stablemarkets/KeilSim documentation built on May 22, 2019, 2:50 p.m.