Harder programming questions do a worse job of predicting outcomes