Online Program

Identifying multiple regulation

*Denis Agniel, Harvard University 
Tianxi Cai, Harvard University 

Keywords: multiple outcomes, sparse regression, hierarchical lasso, closed testing, semi-parametric regression, resampling

How do you decide what interventions will have the most impact on multiple outcomes? Often it is suspected that one can identify a sparse, shared set of predictors for many related outcomes. Then policymakers would want to target that smaller set of predictors for interventions. Furthermore, since not every predictor that is related to the outcomes will necessarily be related to all of them, we would like to identify for each predictor which outcomes it is associated with. In particular, the most efficient policies will target predictors that are important for multiple outcomes, which we will call "multiple regulators". We propose a method for identifying multiple regulation in a large class of models, including semi-parametric regression models, based on the hierarchical lasso (Zhou and Zhu, 2010). We further offer a resampling method to assess the variability in our estimator. And we finally propose a closed testing procedure to assess multiple regulation in the presence of randomness in the observed data.