What is a relevant control?: an algorithmic proposal

Abstract

Individualized inference (or prediction) is an approach to data analysis that is increasingly relevant thanks to the availability of large datasets. In this paper, we present an algorithm that starts by detecting the relevant observations for a given query. Further refinement of that subsample is obtained by selecting the ones with the largest Shapley values. The probability distribution over this selection allows to generate synthetic controls, which in turn can be used to generate a robust inference (or prediction). Data collected from repeating this procedure for different queries provides a deeper understanding of the general process that generates the data.