I think another factor is that the spreadsheet model has basically zero barrier to entry from a users perspective. You can have elementary school students punch data from a classroom experiment into an excel table, which is great!
Any replacement system which, for example, enforced a strong separation between operations, input reference data and output result data would require users to learn the model before attempting to use the software. This is a pretty big ask, especially since lots of small-scale users wouldn't see an immediate benefit. I think of it like the tradeoff between dynamic and static typing when programming- it's the same "upfront mental overhead versus long term maintainability" question IMO.
Because there is no better alternative (yet)?
A better alternative needs to be really better, to justify the effort of people relearning how to do things in this better tool then.