Understand the workload.

Take the workload difference between in house versus public usage.

If one releases software into the public domain then one has no reasonable expectation as to the kinds of data driving your code. For this kind of general purpose, heavily reused software then code coverage with, unit testing and mocking makes the most sense is my experience.

However, one can also write code for a very specific, limited usage. In those cases I find regression testing with regression data works best. They are not paying me to find bugs to satisfy code coverage.

I find understanding the workload is a significant factor in determining the kinds of testing I use.