Many of these old systems rarely require changes in most modules. (As in, of the hundreds of thousands of lines in the system, many subsections just never get touched).
This means for the the parts that never get touched, there is a negative return on investment for rewriting them, so there has to be a correspondingly larger positive return on investment in the higher churn parts to make a rewrite worthwhile. Most of the times when this is studied, the return on investment from the rewrite is either tiny or net negative from this effect.
Ah you say, then just rewrite only the parts that tend to change a lot, and leave the other stuff as COBOL. That sounds good right? Except that generally the cross language interop code needed becomes an absolute nightmare. I've dealt with this in an near ideal enviroment, and it was still a royal pain.
This near ideal environment was: COBOL code restricted to just a subset used by the company, which was being translated to C code by the companies own compiler, which allowed for a C-compatible ABI to be defined allowing for surprisingly simple and direct C to Cobol interop. Yet even then the code code was still a royal pain, just due to how COBOL data structures differ so much from the idiomatic data structures for C or C++ code. My code was actually C# code interacting via C++/CLI with some C++ code that interacted with the the COBOL. But the interop code was still some of the worst code to touch.