Paper Repository for Cyclic Permutation Test (CPT)

This repository contains the code to implement all examples in our paper: An Assumption-Free Exact Test For Fixed-Design Linear Models With Exchangeable Errors. The folder code/ contains all R files for implementation and bash files for submitting the jobs to a cluster with Slurm system. The total computation load is ~20000 CPU hours to reproduce all experimental results in this paper. I used 256 cores for ~7 days.

• CPT.R contains the implementation of cyclic permutation test. It depends on the "gaoptim" package for the genetic algorithm which has been removed from CRAN on 2018-06-17. But it has better performance than other CRAN packages for our purpose so we keep it. The following code can be used to download the package.

devtools::install_url("https://cran.r-project.org/src/contrib/Archive/gaoptim/gaoptim_1.1.tar.gz")

• CPT_expr_prepareX.R contains the template to search for good ordering for CPT using Genetic Algorithm. It takes four external inputs: Xdist for the type of design matrices, ratio for the ratio between n and p, ninds for the number of indices to be tested, seed for random seed. The output includes the desing matrix (X), a weak ordering by running the genetic algorithm with 1000 samples (ordering1) and a strong ordering by running the genetic algorithm with 10000 samples (ordering). To reproduce the matrices used in the paper, run the following code on Shell.

cd CPT/code
./gen_jobs_prepareX.sh params_prepareX.txt

Each row in the file params_prepareX.txt contains the four external inputs, namely Xdist, ratio, ninds and seed, corresponding to an experiment. The bash file gen_jobs_prepareX.sh will automatically generate bash files to call CPT_expr_prepareX.R in results/ folder for each row of params_prepareX.txt, submit each to a core, and save the output in the data/ folder. If the user wants to reduce the number of experiments or add other experiments, the simplest way is to create another .txt file mimicing params_prepareX.txt and including the desired experimental setups and run ./gen_jobs_prepareX.sh NEW_TXT_FILE.txt.

• CPT_expr.R contains the code template for all numerical experiments in Section 3 and Appendix A. It takes five external inputs: Xdist for the type of design matrices, epsdist for the error distribution, ratio for the ratio between n and p, ninds for the number of indices to be tested, seed for random seed. For given inputs, it required CPT_expr_prepareX.R to be ran at least once with Xdist, ratio, ninds and seed so that its output is stored in data/ folder; otherwise the job will be halted. The output of CPT_expr.R is a data frame containing the inputs as well as the relative signal strength and the power for each method. To reproduce the results in Section 3 and Appendix A, first run the following code on Shell.

cd CPT/code
./gen_jobs_CPT.sh params_CPT_expr.txt

Each row in the file params_CPT_expr.txt contains the five external inputs, namely Xdist, epsdist, ratio, ninds and seed, corresponding to an experiment. The bash file gen_jobs_CPT.sh will automatically generate bash files to call CPT_expr.R in results/ folder for each row of params_CPT_expr.txt, submit each to a core, and save the output in the data/ folder. If the user wants to reduce the number of experiments or add other experiments, the simplest way is to create another .txt file mimicing params_CPT_expr.txt and including the desired experimental setups and run ./gen_jobs_CPT.sh NEW_TXT_FILE.txt. After the jobs are done, run the following code to aggregate the results from each core.

R CMD BATCH agglomerate.R

This will generate a file power_CPT_expr.RData in data/ folder. It is currently included for users who do not want to reproduce the experiments.

• CPT_GA_SS.R contains the code template to compare genetic algorithm with stochastic search as in Figure 2. It takes four external inputs: Xdist for the type of design matrices, algo for the algorithm (GA or SS), popSize for the population size of GA (not matter for SS), and seed for random seed. The output of CPT_GA_SS.R is a vector of O* values. To reproduce the results in Figure 2, first run the following code on Shell.

cd CPT/code
./gen_jobs_GA_SS.sh params_GA_SS.txt

Each row in the file params_GA_SS.txt contains the four external inputs, namely Xdist, algo, popSize and seed, corresponding to an experiment. The bash file gen_jobs_GA_SS.sh will automatically generate bash files to call CPT_GA_SS.R in results/ folder for each row of params_GA_SS.txt, submit each to a core, and save the output in the data/ folder. If the user wants to reduce the number of experiments or add other experiments, the simplest way is to create another .txt file mimicing params_GA_SS.txt and including the desired experimental setups and run ./gen_jobs_GA_SS.sh NEW_TXT_FILE.txt. After the jobs are done, run the following code to aggregate the results from each core.

R CMD BATCH agglomerate.R

This will generate a file CPT_GA_SS.RData in data/ folder. It is currently included for users who do not want to reproduce the experiments.

• illustrate_perm.R produces Figure 1.
• CPT_GA_plot.R produces Figure 2.
• CPT_plot.R produces all Figure 3-16.
• expr_functions.R contains all helper functions.