We study deep neural networks and their use in semiparametric inference. We establish novel nonasymptotic high probability bounds for deep feedforward neural nets. These deliver rates of convergence that are sufficiently fast (in some cases minimax optimal) to allow us to establish valid second‐step inference after first‐step estimation with deep learning, a result also new to the literature. Our nonasymptotic high probability bounds, and the subsequent semiparametric inference, treat the current standard architecture: fully connected feedforward neural networks (multilayer perceptrons), with the now‐common rectified linear unit activation function, unbounded weights, and a depth explicitly diverging with the sample size. We discuss other architectures as well, including fixed‐width, very deep networks. We establish the nonasymptotic bounds for these deep nets for a general class of nonparametric regression‐type loss functions, which includes as special cases least squares, logistic regression, and other generalized linear models. We then apply our theory to develop semiparametric inference, focusing on causal parameters for concreteness, and demonstrate the effectiveness of deep learning with an empirical application to direct mail marketing.

This paper proposes a procedure for assessing the sensitivity of inferential conclusions for functionals of sparse high-dimensional models following model selection. The proposed procedure is called targeted undersmoothing. Functionals considered include dense functionals that may depend on many or all elements of the high-dimensional parameter vector. The sensitivity analysis is based on systematic enlargements of an initially selected model. By varying the enlargements, one can conduct sensitivity analysis about the strength of empirical conclusions to model selection mistakes. We illustrate the procedure's performance through simulation experiments and two empirical examples.

We study the welfare implications of personalized pricing implemented with machine learning. We use data from a randomized controlled pricing field experiment to construct personalized prices and validate these in the field. We find that unexercised market power increases profit by 55%. Personalization improves expected profits by an additional 19% and by 86% relative to the nonoptimized price. While total consumer surplus declines under personalized pricing, over 60% of consumers benefit from personalization. Under some inequity-averse welfare functions, consumer welfare may even increase. Simulations reveal a nonmonotonic relationship between the granularity of data and consumer surplus under personalization.