This repository reproduces the results reported in arXiv version 1 the following paper:

Z. Niu, A. Chakraborty, O. Dukes, E. Katsevich. "Reconciling model-X and doubly robust approaches to conditional independence testing." (arXiv)

The code to reproduce the numerical simulations is in simulation-code, as well as the accompanying R package symcrt. Some of the simulations are based on the simulatr R package and associated Nextflow pipeline. The simulatr specifier objects are stored in simulation-code/sim_spec_objects. The results outputted by the simulation code are stored in simulation-results. These results are split into three parts: the primary simulations of the paper (underlying all figures but Figures 2, 5 and 14-16), the simulation to assess the confounding levels of previous studies (underlying Figure 2), the simulation to compare the estimation performances of the lasso and post lasso (underlying Figure 5), the simulation to compare the Type-I error of GCM to that of normalized dCRT for sparse outcome model (underlying Figure 14), and the simulation to illustrate how asymmetry of dCRT can affect the Type-I error (underlying Figure 15 and 16). The results for these parts are stored in subdirectories benchmarking, confounding, diagnostic, GCM_vs_dCRT and asymmetry_investigation of simulation-results, respectively. Code to produce figures based on these results is stored in plotting-code, and the corresponding figures are stored in figures. The manuscript LaTeX source is named as symcrt-manuscript-aos.tex.

• Identify a computing environment (e.g. a laptop, high performance computing cluster, or cloud computing platform) you would like to use. If you would like to reproduce a subset of Figures 3-13, a distributed computing environment (i.e. cluster or cloud) is strongly recommended.
• Check that you have R version 4.1.3 or higher installed. Copies of all necessary R packages will be installed for you via renv.
• Install Nextflow or upgrade to the latest version, and configure it for use in your computing environment.
• Clone this repository into your computing environment.
• Optional: Determine how much RAM is available to each core (code defaults to 8GB).
• Optional/advanced: Determine which Nextflow profile you would like to use (code defaults to standard).
• Delete any pre-computed objects you would like to re-compute, including simulatr specifier objects under simulation-code/sim_spec_objects, simulation results under simulation-results, or figures under figures.
• Open a shell and navigate to the symcrt-manuscript directory.
• Launch the computations via bash run_all.sh <MAXGB> <PROFILE>, where <MAXGB> should be replaced by the available RAM per core in GB (or omitted if you would like to use the default of 8GB) and <PROFILE> should be replaced by the Nextflow profile (or omitted if not using profiles). This will re-compute (only) the deleted pre-computed objects. If no pre-computed objects are deleted, then the code will not do anything.

ZN was partially supported by the grant "Statistical Software for Single Cell CRISPR Screens" awarded to EK by Analytics at Wharton. OD was partially supported by FWO grant 1222522N and NIH grant AG065276. EK was partially supported by NSF DMS-2113072 and NSF DMS-2310654. We acknowledge help from Timothy Barry with our simulation studies and the underlying computational infrastructure, including his simulatr R package and Nextflow pipeline. We acknowledge dedicated support from the staff at the Wharton High Performance Computing Cluster. We acknowledge Lucas Janson for providing details about the simulation setting in Candes et al (2018). We acknowledge Eric Tchetgen Tchetgen for helpful discussions on hypothesis testing in the semiparametric models.