Does language-as-used fit a self-paced reading paradigm ? ( The answer may well depend on how you model the data . )