We investigate past-tense formation preferences for five irregular English verb classes. We gathered data on a large scale using a nonce probe study implemented on Amazon Mechanical Turk. We compare a Minimal Generalization Learner (which infers stochastic rules) with a Generalized Context Model (which evaluates new items via analogy with existing items) as models of participant choices. Overall, the GCM is a better predictor, but the the MGL provides some additional predictive power. Because variation across speakers is greater than variation across items, we also explore individual-level factors as predictors. Females exhibited significantly more categorical choices than males, a finding that can be related to results in sociolinguistics.