> And you always have to make a choice since utility functions are unobservable. If the utility function isn't provided to you directly, then you should infer

> And you always have to make a choice since utility functions are unobservable. If the utility function isn't provided to you directly, then you should infer the functional form from the data with a reasonable prior, and use sampling techniques to show the uncertainty in your answer due to uncertainty in the utility function. There's nothing wrong with utility functions as a model sense they are exactly dual to any 'reasonable' choice function, meaning the space they model is the same. If you find them mathematically inconvenient, then you can use choice functions directly, but raw choice functions are usually not very helpful for modeling or predicting anything at all.