Replication code for Precise Unbiased Estimation in Randomized Experiments using Auxiliary Observational Data

This contains the code to reproduce the results in the manuscript Precise Unbiased Estimation in Randomized Experiments using Auxiliary Observational Data (Gagnon-Bartsch and Sales, et al).

Obtaining the code and data

• Although not required, the code here is intended to be run in Docker.
  • A docker image containing this code can be found at https://osf.io/d9ujq/.
  • To use the docker image, first download it, then install it and run it with
  • docker load precise2023.tar.gz
  • docker run -it --rm -p 127.0.0.1:80:80 precise2023
  • For reference, the Dockerfile is in the docker directory.
• The data files are not included here.
  • Data files are available at https://osf.io/j6esa/.
  • Note that the data can be downloaded, but not redistributed.
  • Likewise, the data is not included in the docker image.
  • The data is automatically downloaded with data/download.Rmd

Notes on the code

All of the analyses can be run by typing make in the root directory. (This also first downloads the data.)

• The analysis folder contains code notebooks in markdown format that can be run, e.g., in Jupyter.
  • 01_remnant_model_1.md: Fits the auxiliary data model for the first 22 experiments.
  • 02_remnant_model_2.md: Fits the auxiliary data model for the other 11 experiments.
  • 03_assemble_data.Rmd: Assembles the full 33-experiment dataset, including the predictions from the models above.
  • 04_estimate.Rmd: Runs the main analyses of the paper and saves the output.
  • 05_outputs.Rmd: Produces the plots, tables, and other numerical results that are in the paper.
  • 06_appendixD.Rmd: Reproduces the figures in Appendix D.
  • scripts: Additional python scripts used by the remnant models.
• The data/original folder should contain the raw data downloaded from https://osf.io/j6esa/.
• The docker folder only contains files necessary to build the docker image.
• The output folder contains the output (figures, tables, etc.) from running the code.
• The package folder contains an R package that implements the statistical methods from the paper.