Statistical Analysis and Design of Crowdsourcing Applications

This thesis develops methods for the analysis and design of crowdsourced experiments and crowdsourced labeling tasks. Much of this document focuses on applications including running natural field experiments, estimating the number of objects in images and collecting labels for word sense disambiguation. Observed shortcomings of the crowdsourced experiments inspired the development of methodology for running more powerful experiments via matching on-the-fly. Using the label data to estimate response functions inspired work on non-parametric function estimation using Bayesian Additive Regression Trees (BART). This work then inspired extensions to BART such as incorporation of missing data as well as a user-friendly R package.