You can't without defining yet another field, usually. It's really annoying to double the number of columns so that [time_started, time_finished] becomes [time_started_exists, time_started, time_finished_exists, time_finished].
NULL values makes business logic far more compact and intuitive. For enumerated values in fields it's easy enough to define another value in the same field to mean 'unknown' or 'not entered', but you can't do that for strings, numbers, datetimes, etc. -- you have to throw in a bunch of unwieldy additional boolean fields instead.
The underlying problem here is that SQL lacks Sum Types (aka Tagged Unions). Such types solve all these problems effortlessly.
In contrast to what SQL has, Sum Types combined with Product Types (which is basically what a row is) are actually a universal way to model all possible data[0]. (Of course you may want syntax sugar, etc. on top of that, but Sums and Products at the bottom is sufficient.)
[0] I'm actually not sure if I should qualify that -- I believe Sums+Products can actually model anything, assuming you allow recursive type definitions -- which might be hard to make perform well. Storing a linked list in a database field, e.g. might not be the best idea.
NULL values makes business logic far more compact and intuitive. For enumerated values in fields it's easy enough to define another value in the same field to mean 'unknown' or 'not entered', but you can't do that for strings, numbers, datetimes, etc. -- you have to throw in a bunch of unwieldy additional boolean fields instead.