Yes, but some of the possible compression functions approximate a random function more closely than others. If we have information that a given construction does not, then we may be obligated not to approximate it as such.
I did some Monte Carlo testing and learned some things :-). Results and code are posted here:
* Mixing the input chaining value back to the output (D-M) causes the result to more closely approximate a random function (for better or for worse). So MD4-SHA2 do, in fact, experience the entropy reduction effect.
* D-M does convert the worst-case of a completely degenerate compression function (e.g., one which returns the same image every time) into a best case of no entropy loss.
* A pure M-D construction using a random permutation (say AES) for the CF is made worse by randomly mixing into the output.