Improving statistical methods enhances the productivity of researchers, helps ensure correctness of results, and unlocks more value from our data. We believe it is incredibly important to invest in our fundamental capabilities to use data effectively.
In support of academic work that addresses methodological needs and produces generalizable knowledge for the broader research community, Facebook is pleased to announce the winning proposals of our Statistics for Improving Insights and Decisions request for proposals:
Fast multilevel regression trees for heterogeneous effect estimation
Paul Richard Hahn, Arizona State University
Kernel Population Weighting for Survey Nonresponse
Erin Hartman, UCLA
DeepForecast: Leveraging forecasts on large scales of related time series
Rob J Hyndman, Monash University
All three winning proposals aim to provide researchers with practical tools to produce better estimation and improve data-driven decision making within organizations like Facebook. No Facebook data will be provided to award recipients and research will be conducted independently of Facebook.
Richer, larger data sets hold the potential for understanding treatment effect heterogeneity in randomized experiments and in observational data — allowing researchers to understand when and how treatments are effective or harmful. Bayesian Additive Regression Trees (BART) are one of the top performing methods for modeling effect heterogeneity, but researchers have faced challenges scaling it to larger data sets and accounting for hierarchical structure in data sets. “Fast multilevel regression trees for heterogeneous effect estimation” will build on promising work on Bayesian Causal Forests (BCF) and implement key improvements to the BART algorithm that allow it to scale to much larger data sets.
Whether they are produced by surveys or other opt-in mechanisms, many modern data sets exhibit substantial non-response bias that makes it challenging for researchers to draw the conclusions they would like. “Kernel Population Weighting for Survey Nonresponse” helps researchers capitalize on the availability of large amounts of auxiliary data available in many settings to help mitigate non-response bias by efficiently and generically estimating weights that ensure that biased samples are as representative of the intended population as possible.
The majority of practical forecasting techniques treat each time series as an independent sequences of observations, despite the fact that in many applications they may share a large amount of common structure. For instance, our open source forecasting package Prophet has to re-learn seasonality and growth on each new data set, despite that many of the time series that we study exhibit very similar patterns. “DeepForecast: Leveraging forecasts on large scales of related time series” plans to develop a procedure that can flexibly fit models that simultaneously model set of related time series, borrowing strength across observations of sequences and improving forecasts. This is a promising path forward toward better forecasting performance and builds on rapidly developing deep learning research and software.
We received 78 proposals in response to our RFP and it was incredibly difficult for our review team to choose just three winners. We would like to acknowledge the following proposals for being particularly exciting areas of future research:
Robust Pure-Exploration and Experimentation with Environmental Context
Kevin Jamieson, University of Washington
Toward robust causal inference: study design and design-based inference
Peng Ding, University of California, Berkeley
GProphet: Analyst-in-the-loop Forecasting using Gaussian Processes
Hong Ge, University of Cambridge
Laplace approximation for non-response adjustments with large surveys
Andrew Gelman, Columbia University
Efficient & Compact Embedding Summaries for Time-varying Networks
Danai Koutra, University of Michigan
Modelling Social Processes on Time-varying Multi-layered Networks
Matthew Nunes, University of Bath
Thank you to all the researchers that submitted proposals, and congratulations to the winners. To view our currently open research awards and to subscribe to our Research Awards email list, visit our Research Awards page.