Title: Minimax learning for average regression functionals with an application to electoral accountability and corruption
Authors: Chen Qiu - London School of Economics (United Kingdom) [presenting]
Abstract: A new minimax methodology is proposed to estimate average regression functionals, which cover many empirical problems including average treatment effect. Featured in penalized series space, this strategy exploits minimax property of a vital nonparametric component of average regression functional and aims to directly control key remainder bias. We then construct a new class of estimators, called minimax learners, and study their asymptotic properties when number of controls over sample size goes to zero, constant and infinity, respectively. Root-n normality is established under weak conditions for all three cases. Minimax learners are fast to implement due to their minimum distance representation. In simulations where selection bias is mild, they behave more stably, show less mean square error and do not over control. When applied to a previous work that studies effect of electoral accountability on corruption, minimax learners behave less erratically and lead to more coherent conclusion, even when number of controls becomes very large.