This paper describes our approach to the CLPsych 2021 Shared Task, in which we aimed to predict suicide attempts based on Twitter feed data. We addressed this challenge by emphasizing reliance on prior domain knowledge. We engineered novel theory-driven features, and integrated prior knowledge with empirical evidence in a principled manner using Bayesian modeling. While this theory-guided approach increases bias and lowers accuracy on the training set, it was successful in preventing over-fitting. The models provided reasonable classification accuracy on unseen test data (0.68<=AUC<= 0.84). Our approach may be particularly useful in prediction tasks trained on a relatively small data set.