Semiparametric panel data models using neural networks

This is a past event

99 people went

Details

Agenda:
6:30 - 7:00 food, drinks, meet people
7:00 - 7:15 Intros and announcements
7:15 - 8:30ish Presentation, Q&A
Off to data drinks

For event location please register here:

https://nvite.com/Semiparametricpanel/q2v22v

Presentation:

Recent years have seen truly impressive advances in machine learning, largely in predictive arenas such as computer vision and natural language processing. Incorporation of these advances into the rest of statistical practice has been slow, but is gaining steam. The blending of classical causal inference with ML-derived predictive techniques is a new and booming research area. It promises to allow researchers not only to predict outcomes, but to predict heterogeneous consequences of actions for individuals with different characteristics, allowing decisionmakers to target their actions to the areas where they'll have the greatest effect.

This talk will give a broad overview of ML for heterogeneous causal effects, before taking a dive into early work on a semiparametric modeling approach built around (deep) neural nets (https://arxiv.org/abs/1702.06512). By combining parametric structure, neural nets, and inferential principles from the literature on smoothing splines, these models allow powerful prediction of deeply nonlinear systems with certain forms of nonstationarity, unobserved cross-sectional heterogeneity, and potentially-known partial parametric structure. They also give provide confidence intervals with as-advertised coverage properties, which allows for bounds on prediction as well as inference on marginal effects.

This is very much a work in progress. The goals of the talk are (1) to seek feedback/get ideas, and (2) to advertise the new R package panelNNET, which implements all of this.

Speaker Bio:

Andrew Crane-Droesch works as an economist for the USDA, where he focuses on ML tools and "big data" for crop yield forecasting and climate impacts research. Prior to that, he was the "Massive Data Institute" postdoctoral fellow at Georgetown's McCourt School of Public Policy. His PhD is from UC Berkeley's Energy & Resources Group.

Read the research paper: https://arxiv.org/abs/1702.06512