PS-PPI

This repository is the official repository of the paper "A Unified Framework for Inference with General Missingness Patterns and Machine Learning Imputation".

Prerequisites

Environment Setup

The simulation code is based on Python. However, if you would like to run simulations with multiple imputation, you need to set up both a Python and R environment. If you are not familiar with R or Python/Conda, please refer to their documentations as follows:

• R Installation Guide
• Conda Installation Guide

To setup the environments for Python, we recommend using conda to create an environment:

conda env create -f environment.yml
conda activate ps-ppi

For R, ensure that you have the 3.16.0 verison of the mice package installed. You can do this by running the following commands in R.

install.packages("mice")

Other necessary packages

If you want to run the simulations with the official implementation of the PPI++ methods, you should install the ppi_py package thorugh here. We also reimplemented the PPI++ methods based on the framework of PS-PPI, which is the method called PPI++_reimp in the code.

Running Simulation

We provide the following example command to showcase how to run the proposed PS-PPI method on the linear regression settings:

python general_missing_linear_regression.py \
  --use_estimated_ps True \
  --prediction_noise 0.2 \
  --prediction_bias 0.1 \
  --methods PS-PPI

By running the above command, the script will run a simulation study with the specified settings and output the results into the results/ directory.

Key CLI Options

• --methods – comma-separated list of estimators to evaluate (WCCA, PPI++, PS-PPI, CCA, MI, PPI++_reimp, PS-PPI_closedform). Note that PS-PPI_closedform is a variant of PS-PPI that uses closed-form imputation variance estiamtion for linear regression to speed up the computation.
• --n_simulations – number of repeated simulations.
• --prediction_noise / --prediction_bias – perturbations injected into the surrogate predictors.
• --use_estimated_ps – enable propensity score estimation inside the simulator; paired with --mispecified_ps_type for misspecification experiments.
• --missing_mechanism – missingness type (MAR or MCAR).

Run python general_missing_linear_regression.py --help for the full list of command-line flags.

Batch Job Submission

To enable running simulation in batch through a job scheduler (e.g., SLURM), we provide example shell scripts submit_job_general_missing_linear_regression.sh and submit_job_general_missing_linear_regression_mispecified_ps.sh. You can modify the parameters inside these scripts to run the simulations with different settings as needed.