Home
Monday, 13 July, 2026
Maksim Smirnov: Essays on Instrumental Variables
Dissertation Committee:
Stanislav Anatolyev (CERGE-EI, chair)
Nikolas Mittag (CERGE-EI)
Vasily Korovkin (Universitat Pompeu Fabra)
Defense Committee:
Byeongju Jeong (CERGE-EI, chair)
Achim Ahrens (CERGE-EI)
Arnošt Komárek (Charles University, MFF)
Referees:
Kirill Borusyak (UC Berkeley)
Laurent Davezies (Institut Polytechnique de Paris)
Meeting link:
https://cerge-ei.webex.com/cerge-ei/j.php?MTID=m9d12d985df2638e89d71f1e0e60b8e82
Meeting number:
2743 118 3108
Meeting password:
283161
Abstract:
In this thesis, I study instrumental variable models and propose new methods for their analysis. I derive novel theoretical results, based on which I propose new statistical procedures that are relevant for applied research. In addition to analytical derivations, I run simulations and apply the proposed methods to relevant empirical contexts. In Chapter 1, I consider a nonparametric instrumental variable model with binary treatment and propose a simple test of unobserved heterogeneity of treatment effects. The test is built on a novel statistical result that allows for the identification of the average second regression derivative by quadratic least-squares projections. Based on this statistical result, I propose a test of heterogeneity of average treatment effects across margins of participation, using the marginal treatment effects (MTE) methodology. In particular, I exploit the link between the slope of the MTE curve and the second derivative of an unknown regression function. The test is easy to implement and can be run using off-the-shelf software. Further, I illustrate the performance of my test with simulations based on an influential empirical setup. Finally, I propose a version of my test that checks for a monotone MTE curve and, more generally, monotone regression derivatives. Both versions of my test have implications for welfare maximization and policy analysis. In Chapter 2, we study the setting with many weak instrumental variables and clustered dependence. We propose a new estimator of structural parameters, which is simple in implementation. The estimator combines leave-cluster-out methodology with inverse-root-cluster-size reweighting, which together provide robustness and desirable asymptotic properties. We also propose an asymptotic variance estimator, which allows for valid inferences based on pivotized test statistics. Finally, we show the properties of our procedure using simulations and real data from an influential study.
Full Text: "Essays on Instrumental Variables"
