Software
rswjax
A JAX implementation of optimal representative sample weighting, drawing heavily on the original rsw implementation of Barratt, Angeris, and Boyd (rsw). Thanks to rewriting some core operations in JAX, it is significantly faster than rsw for medium-large datasets, especially those with many columns (ex: 5k+ rows, 20+ columns). In addition, it adds a number of quality of life improvements, like early stopping if optimization produces NaNs, warnings for common issues like not including loss functions for every column, and a broader test suite. I started this project because of my interest in regularized raking, and wanting to iterate faster on large weighting problems.
retrodesign
retrodesign provides tools for working with Type S (Sign) and Type M (Magnitude) errors, as proposed in Gelman and Tuerlinckx (2000) and Gelman & Carlin (2014). In addition to simply calculating the probability of Type S/M error, the package includes functions for calculating these errors across a variety of effect sizes for comparison, and recommended sample size given “tolerances” for Type S/M errors. To improve the speed of these calculations, closed forms solutions for the probability of a Type S/M error from Lu, Qiu, and Deng (2018) are implemented. The broader goal of this project was to make it easier for researchers to understand these issues in their work, and it’s gratifying the package has been able to do that.