Welcome to the PrivLCM package. Here, we implement the approach of Nixon et. al (2022).
Here, we provide the code to fit our differentially-private Bayesian latent class model for synthetic data creation or posterior inference. The main steps of our approach are:
- Subselect a set of marginal counts and add noise using the Geometric mechanism for differential privacy.
- Post-process these counts to be able to estimate probabilities for any table cell. This is done using ideas from composite likelihood and the Dirichlet Process Mixture of Product Multinomials model.
Please see the small vignette (example.Rmd) to get started!