Tetsuya Kaji

Assistant Professor of Econometrics and Statistics
University of Chicago Booth School of Business

PhD, Economics and Statistics, MIT, 2018

C.V.



Research Interests

Weak Identification, Empirical Processes, Semiparametric Theory, Program Evaluation, Financial Risk Measures, Machine Learning


Working Papers

Switching to the New Norm: From Heuristics to Formal Tests using Integrable Empirical Processes

Asymptotics of L-statistics can be tricky in some applications. I develop a new way to characterize them using integrable empirical processes and functional delta methods. Bootstrap validity is also shown. Applications include outlier robustness analyses and Lorenz dominance tests.

Theory of Weak Identification in Semiparametric Models

Weak identification is analyzed in general semiparametric models. It is discovered that many insights from the weak IV literature (though not all) carry over to arbitrary weakly identified models: impossibility of consistent estimation, existence of "reduced-form parameters," etc. Then a novel notion of weak efficiency is defined. Simulation improves 2SLS and GMM in weak IV. (Yes, even oracle GMM is not "efficient" in weak IV!)

Controlling Tail Risk Measures with Estimation Error (with H. Kang)

Financial risk control inevitably involves estimation of risk. If risk intends to control the probability of bad events, we construct a way to control the "true risk" exploiting the knowledge of estimation error. Such risk measures, named tail risk measures, include Value-at-Risk and expected shortfall. An empirical application controls expected shortfall in portfolio management.

Assessing Outcome-Dependent Heterogeneity in Treatment Effects (with E. Manresa)

Treatment heterogeneity is crucial in policy targeting. First, we interpret the popular heterogeneity measure, the quantile treatment effect, by the principle of equal effects. Second, we relax it to the principle of least effects and propose bounds on subgroup treatment effects. Third, we provide sharp second-order stochastic dominance bounds on the distribution of individual treatment effects.

Quantile Regression in 𝑳1 (with V. Chernozhukov)

We attempt to generalize the L1 convergence of empirical quantile functions derived in "Switching to the New Norm..." to quantile regression estimators. We obtained partial results.

Deep Inference: Artificial Intelligence for Structural Estimation (with E. Manresa and G. Pouliot)

This paper is our first step to answer "When should we (economists) use deep neural networks and why?" We apply neural network classifier techniques to otherwise intractable structural estimation problems.


Publications

Extremal Quantile Regression (with V. Chernozhukov and I. Fernández-Val)

Handbook of Quantile Regression, ed. by R. Koenker, V. Chernozhukov, X. He, and L. Peng, Chapman & Hall/CRC, 2017.