Mind the data gap: Missingness Still Shapes Large Language Model Prognoses