On the Efficacy of Adversarial Data Collection for Question Answering: Results from a Large-Scale Randomized Study