A missing value tour from predictive models to causal inference

par Julie Josse (CMAP, École Polytechnique, Saclay)

mardi 26 févr. 2019, 10:00 → 12:00 Europe/Paris

112 (Braconnier)

This presentation will provide an update on how to manage missing data (imputation, mechanism that generate missing values, etc). Most of the methods have been developed with the objective of estimating the parameters and their variance as best as possible with missing values and not in a predictive framework. Thus, many practical questions have not been studied much: what to do with missing data in a test dataset, should the response variable be integrated into imputation methods.... In particular, we will focus on establishing predictive models with missing data with random forests that have the advantage of being able to be used to make causal inferences with double robust  methods.