Data wrangling, computational burden, automation, robustness and accuracy in ecological inference forecasting of R×C tables

Abstract

This paper assesses the two current major alternatives for ecological inference, based on a multinomial-Dirichlet Bayesian model and on mathematical programming. Their performance is evaluated in a database made up of almost 2000 real datasets for which the actual cross-distributions are known. The analysis reveals both approaches as complementarity, each one of them performing better in a different area of the simplex space, although with Bayesian solutions deteriorating when the amount of information is scarce. After offering some guidelines regarding the appropriate contexts for employing each one of the algorithms, we conclude with some ideas for exploiting their complementarities.