Repository for the paper “Survey calibration for causal inference: a simple method to balance covariate distributions”

Paper

Here is the first version of the paper: “Survey calibration for causal inference: a simple method to balance covariate distributions” (2023.10.14, arxiv version)

Acknowledgements

This work was financed by the National Science Centre in Poland, OPUS 22 grant no. 2020/39/B/HS4/00941.

Reproduction of the results

Setup

Install relevant packages for the paper. The jointCalib package is available at CRAN but we install the package from github (development version). The IPS package used in the Sant’Anna et al. (2022) and the kbal package used in the Hazlett (2020) are available only on github.

UPDATE: as of 2023.10.19 (version: 0.14.2.9001) the WeightIt package implements methods proposed in my paper (yay!). See WeightIt NEWS.md file and the documentation of ebal, npcbps, optweight and energy. Importantly, Noah Greifer implementation allows for multinomial treatment studies.

install.packages(c("remotes", "ebal", "mvnfast", "data.table", "ggplot2", "laeken", "xtable", "glue", "stringr"))
remotes::install_github("ncn-foreigners/jointCalib@dev")
remotes::install_github("chadhazlett/KBAL")
remotes::install_github("pedrohcgs/IPS") 
remotes::install_github("ngreifer/WeightIt") ## 0.14.2.9001

Notebooks and results

• Notebooks:
  • simulation 1 (EB),
  • simulation 2 (CBPS)
  • processing results for the paper
  • simulation 3 (why it works)
• Simulation results may be found in folder results/
• Tutorial for the jointCalib package
• A minimal example using the WeightIt package.
• Codes for the proposed method in Stata and Python [work in progress].

Highlights

Results for the distributional entropy balancing (DEB)

Results for the distributional propensity score (DPS)

Why it works?

Including quartiles ($A_q$) and deciles ($A_d$) along with X approximates the relationship between Y and X by local linear functions (as in segmented regression) and thus improves the estimates, as EB and CBPS assume a linear relationship between Y and calibration/balance variables.