Differential Privacy Applications to Bayesian and Linear Mixed Model Estimation
Abowd, John M.; Schneider, Matthew J.; Vilhuber, Lars
We consider a particular maximum likelihood estimator (MLE) and a computationally-intensive Bayesian method for differentially private estimation of the linear mixed-effects model (LMM) with normal random errors. The LMM is important because it is used in small area estimation and detailed industry tabulations that present significant challenges for confidentiality protection of the underlying data. The differentially private MLE performs well compared to the regular MLE, and deteriorates as the protection increases for a problem in which the small-area variation is at the county level. More dimensions of random effects are needed to adequately represent the time- dimension of the data, and for these cases the differentially private MLE cannot be computed. The direct Bayesian approach for the same model uses an informative, but reasonably diffuse, prior to compute the posterior predictive distribution for the random effects. The differential privacy of this approach is estimated by direct computation of the relevant odds ratios after deleting influential observations according to various criteria.
Differential Privacy; Statistical Disclosure Limitation; Privacy-preserving Datamining; Linear Mixed Models; Quarterly Workforce Indicators; MLE; REML; EBLUP; Posterior Predictive Distribution; Random Effects
Required Publisher Statement: Copyright held by authors.