Fast multilevel regression trees for ...

Project: Research project

Project Details

Description

Fast multilevel regression trees for ... Fast multilevel regression trees for heterogeneous effect estimation Our recent research into the use of Bayesian regression tree models for causal inference, for the purpose of estimating heterogeneous treatment effects from experimental or observational data, has yielded two practical methodological innovations. First, the choice of regularization technique crucial in highly nonlinear regressions can have an outsized impact on the estimates of causal effects (while straight, as opposed to counterfactual, predictions can be relatively unchanged) in the presence of strong confounding (CITE ACIC). To overcome this sensitivity to regularization, we developed a novel model called Bayesian Causal Forests, which gives state-of-the-art estimation performance in data generating regimes with strong confounding (CITE). Second, in the case of experimental data, a modified BCF model can accommodate hierarchical observational units students within classrooms, classrooms within schools, schools within districts, etc. Essentially this takes a multilevel regression model and replaces all of the linear terms with regression forests. To the best of our knowledge, no other method permits similar flexibility in the hierarchical setting. The upshot, relative to unmodified BCF, is uncertainty quantification that is explicitly adjusted for the hierarchical structure of the data. So far, this approach has been applied to a large educational experiment (CITE) and we have plans to apply it to a nutrition experiment in the near future using data from (CITE).
StatusFinished
Effective start/end date5/15/198/31/21

Funding

  • INDUSTRY: Domestic Company: $50,000.00

Fingerprint

Explore the research topics touched on by this project. These labels are generated based on the underlying awards/grants. Together they form a unique fingerprint.