It would need to be researched as to whether or not such an action would produce a collision in practice, but you can consider that a recursive MD5 hash is still only 1 MD5 hash at the end. You are only running the hash on 32 bits of data (the output of your n - 1) instead of on n bits of data. So, considering that the hash algorithm is pseudo-random, you might instead ask whether collisions are more or less likely on 32 bits of data vs on a variable size.
The collisions should be equally unlikely in either case, due to the fact that they are always unlikely with a good algorithm. As to the ability to force a collision, you might defeat known methods, but in principle it should be the same, if your new algorithm was researched.
Finally, you must consider what is the significance of your fear of collisions. As I understand it, collisions allow one to enter instead of the password hashed, another password and it will hash the same. In certificates, it means the ability to generate a new private key for encrypting data (ie. the password of your certificate). So if your login hashes have been compromised, collisions might allow them to enter your site, but not to know the user's passwords exactly. And that will be mainly the cryptographic use of a has algorithm - to hide the password or other source data. In this case, you are better hashing multiple times because it will increase needed CPU to guess the plaintext hashes. You may assume that one your site has been hacked, that all your information is also divulged, and so no great need remains for an attacker to login again to your site with a collided password.
The ability to authenticate a password or other data is reduced by collisions, but again is unlikely in general and so also recursively unlikely as each has as many collisions as on the nth hash only, which is the same as one hash. The subset of data (the result of the n - 1th hash) should not increase the risk of collisions with a pseudo-random hash function. If it is not truely random, the collisions definitely increase, but that does not necessarily make them easier to find or create, though you will loose some iniqueness. With a good hash, this should be minimal and not of any concern over the computational difficulty obtained by recursion.