Publications
We introduce a bias-corrected "two-scale DNN" estimator that linearly combines two distributional nearest neighbors at different subsampling scales. The negative weights that arise from this combination are what let the estimator hit the optimal nonparametric convergence rate under fourth-order smoothness — something standard nearest-neighbor methods cannot achieve. We provide asymptotic normality results and practical inference via jackknife and bootstrap.
This paper proposes a Deep Difference-in-Differences (Deep-DiD) method that combines deep learning with the classic DiD framework to estimate heterogeneous treatment effects. Applied to content creator selection on digital platforms, the method identifies which creators generate the most engagement uplift, enabling data-driven influencer marketing decisions.
We decompose price elasticity estimation into intermediate prediction tasks solved with bagged nearest neighbors and a control-function correction for endogenous prices. Just-in-time compilation and GPU implementations make the resulting pointwise estimator scalable to large datasets. In the final NYC ride-sharing application, we estimate how price sensitivity varies across locations and times using January–February 2024 Uber trip records.
Working Papers
We develop a control function approach for estimating heterogeneous treatment effects over a continuous endogenous treatment — not just binary treatments or covariate-based heterogeneity. Applied to bullet chats (danmaku) on a video platform using a randomized "pre-set bullet chat" instrument, we find a non-monotonic effect: sparse chats reduce viewing, but denser chats increase it, driven by perceived popularity.
Platforms build social ties through rewards and recommendations. Follows produced this way are platform-engineered ties; follows that form on their own are organic ties. Using complete event-level logs from a content platform and a staggered-adoption event study, we find that creators post less after their first engineered follower but more after their first organic one. Engineered followers are less likely to have interacted with the creator before and less likely to receive reciprocation after, so creators respond not only to audience growth but to what a new follower reveals about audience interest.
Consumer response to marketing text depends on how easily messages can be understood. However, existing measures of reading difficulty still rely on surface features, such as average word and sentence length. We introduce word-by-word predictability – estimated with large language models – as a scalable, theory-grounded measure of processing ease.
We embed the full network graph directly into treatment effect estimation using a dual architecture: one neural network models heterogeneous effects as a function of covariates, while a graph convolutional network captures spillovers without imposing assumptions on how interference propagates. This lets the method handle complex interference patterns across diverse network topologies.