Reproduce “Do covariates explain why these groups differ? The choice of reference group can reverse conclusions in the Oaxaca-Blinder decomposition” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Manuel Quintero, Advik Shreekumar, William T. Stephenson, Tamara Broderick
Research from MIT and UC Berkeley demonstrates that selecting different reference groups in the Oaxaca-Blinder decomposition can lead to contradictory conclusions regarding the drivers of group differences, with this instability observed even with advanced machine learning models and large datasets. The work suggests reporting results from both reference groups to prevent misinterpretation.
Key insights
- Sign reversals in the explained and unexplained components of the Oaxaca-Blinder decomposition are not 'measure zero' phenomena; theoretical analysis proves they occur in a substantial portion of the parameter space (up to 50%).
- More complex machine learning models, including state-of-the-art transformer-based foundation models, do not mitigate this issue and can even increase the frequency of contradictory conclusions compared to simpler models.
- The prevalence of sign flips in empirical data can be influenced by specific real-world data structures, such as 'aligned intercept–slope effects' or the presence of unobserved covariates impacting estimated parameters.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.