The debugging I encountered mostly revolved around correcting stack manipulations so that things lined up with the right parameters. The brevity of these languages comes from the stack allowing implicit parametrization; just list your data and then the word you want to use. The downside comes from this property also allowing the stack to leak when under or over-parametrized - enabling mystery behavior or crashes that you wouldn't see with C-style calling conventions.
On the other hand, a major plus: Working with the stack lets you avoid littering your code with named variables. This is what enables Forth words to have extremely high density and readability, and I've incorporated more of this style into my coding elsewhere, so that in the most extreme case, I might go ahead and set up a stack in(for example) an object, and then list out method calls like a high-level Forth operation:
obj.wordA(); obj.wordB(); obj.wordC();
Blending this with more imperative or functional constructs, I can get the best of both worlds: something compact and readable but also self-documenting, with decent type and parameter safeness.
This is one reason that the Forth-like Factor language makes documentation of a word's stack-effect a required part of the syntax of defining a word. I believe these days Factor checks for stack-overflow and -underflow at compile-time, based on the stack-effect declarations.
This is great if you were the one that wrote the DSL and thus understand it intimately (having a prodigious ability to memorize is pretty important too). On the other hand, it sucks if you didn't write the DSL and have to understand how to use it or fix bugs in it because you first have to reverse-engineer the DSL. Reverse-engineering the DSL may take as long as writing it in the first case.
[1] The biggest documentation failure I typically have seen in Forth is a failure to document the word's parameters, i.e. the stack contents going into and returning from words. This is aggravated by the stack manipulation operations duping, swapping, etc. that look more like a game of "15 squares" http://en.wikipedia.org/wiki/Fifteen_puzzle than an algorithm.