After a point, users must develop a deeper understanding of how a given software package works in order to use it effectively.

You mentioned word processing as an example... carrying that along, Microsoft Word faced a huge amount of resistance from WordPerfect users who had internalized WordPerfect's 'stream-of-markup' model for representing formatted text. In both cases, you could highlight text and make it bold, but WordPerfect's model was much clearer to more advanced users. Microsoft has tried to replicate some of this over the years with 'show formatting', but it isn't as effective because the underlying data model is so much more complex. (Where does formatting come from? The character, paragraph, paragraph style... and on and on.)

I should point out that I'm using the term 'data model' vs. 'data structure' deliberately. I would agree with you that users probably don't care about the specific representation for WordPerfect's 'stream of tags'.... only that it's there and has a set of well defined and predictable operations.