Actually the code-length point isn't Paul Graham's. It is Fred Brooks', Paul merely happened to agree with it.
The point appears in The Mythical Man-Month and was based on audits that showed that programmers in different languages were getting the same amount of code written per day in the end, even though that code did very different amounts of stuff.
The cause appeared to be that higher-order languages offer higher-order concepts, and as a result the programmers menally "chunk" at a higher level at first, and then find it easier to debug later because there is less code to understand. For a trivial example, how much do Perl's memory management and hashes reduce the amount of code needed when going from C to Perl, and speed up the programmer?
While it is easy to find plenty of counter-examples (such as your Scheme continuation versus a loop), as a rule of thumb for real projects it really does seem to work.